Email alert
2022 Vol. 44, No. 5
Display Method:
2022, 44(5): 1513-1520.
doi: 10.11999/JEIT211350
Abstract:
In order to improve the efficiency and accuracy of steel quality images detection and promote the automation level of industry, an improved Res-UNet segmentation algorithm is proposed. ResNet50 is used instead of ResNet18 as the encode module to enhance feature extraction capability. Structure like DenseNet is added to encode module, which helps to make full use of shallow feature maps. A new loss function combining weighted Dice loss and weighted Binary Cross Entropy loss (BCEloss) is used to alleviate data imbalance. Data set enhancement strategy ensures that the network learns more features and improves the segmentation accuracy. Compared with the classic UNet, the Dice coefficient of the improved Res-UNet increases by 12.64% and reaches 0.7930. In all, the improved Res-UNet achieves much better accuracy on various defects while requires much less training efforts. The algorithm proposed by this paper is of practical use in the field of steel surface defect segmentation.
In order to improve the efficiency and accuracy of steel quality images detection and promote the automation level of industry, an improved Res-UNet segmentation algorithm is proposed. ResNet50 is used instead of ResNet18 as the encode module to enhance feature extraction capability. Structure like DenseNet is added to encode module, which helps to make full use of shallow feature maps. A new loss function combining weighted Dice loss and weighted Binary Cross Entropy loss (BCEloss) is used to alleviate data imbalance. Data set enhancement strategy ensures that the network learns more features and improves the segmentation accuracy. Compared with the classic UNet, the Dice coefficient of the improved Res-UNet increases by 12.64% and reaches 0.7930. In all, the improved Res-UNet achieves much better accuracy on various defects while requires much less training efforts. The algorithm proposed by this paper is of practical use in the field of steel surface defect segmentation.
2022, 44(5): 1521-1529.
doi: 10.11999/JEIT211416
Abstract:
Due to the lack of mature labeled corpus and large-scale injection molding knowledge graphs for defection diagnosis, industrial knowledge graphs are constructed with high cost and low quality. A framework for constructing industrial knowledge graph based on ontology guidance is developed in this paper. Firstly, the injection molding ontology guided by defect-appearance-cause-scheme chain is designed to limit the collection of web pages. Then, the ontology information is sequentially integrated into the trigger thesaurus to improve the knowledge extraction performance of unstructured web text. Finally, the two-level entity merging method is carried out by combining with the attribute similarity in ontology, which realized the fusion of redundant knowledge. Compared with the existing methods, the accuracy of domain knowledge is higher than 95%, which can be used for tracing the defect quickly.
Due to the lack of mature labeled corpus and large-scale injection molding knowledge graphs for defection diagnosis, industrial knowledge graphs are constructed with high cost and low quality. A framework for constructing industrial knowledge graph based on ontology guidance is developed in this paper. Firstly, the injection molding ontology guided by defect-appearance-cause-scheme chain is designed to limit the collection of web pages. Then, the ontology information is sequentially integrated into the trigger thesaurus to improve the knowledge extraction performance of unstructured web text. Finally, the two-level entity merging method is carried out by combining with the attribute similarity in ontology, which realized the fusion of redundant knowledge. Compared with the existing methods, the accuracy of domain knowledge is higher than 95%, which can be used for tracing the defect quickly.
2022, 44(5): 1530-1538.
doi: 10.11999/JEIT211174
Abstract:
To solve the problem of the slow speed of the existing light field camera calibration algorithm, which makes it unable to quickly calibrate camera parameter changes in industrial inspection and reduce the efficiency of industrial inspection. A fast calibration algorithm for light field cameras is proposed. It selects high-quality, representative sparse views from the light field data based on image clarity, and establishes the sparse light field, which is used for calibration. Compared with the existing optimal approach, the proposed method not only increases the average calibration speed by more than 70%, reducing the average calibration time from 101.27 s to 30.99 s in the existing five datasets, but also maintains the calibration accuracy at the optimal level. The calibration error in the public dataset PlenCalCVPR2013DatasetA is only 0.0714 mm.
To solve the problem of the slow speed of the existing light field camera calibration algorithm, which makes it unable to quickly calibrate camera parameter changes in industrial inspection and reduce the efficiency of industrial inspection. A fast calibration algorithm for light field cameras is proposed. It selects high-quality, representative sparse views from the light field data based on image clarity, and establishes the sparse light field, which is used for calibration. Compared with the existing optimal approach, the proposed method not only increases the average calibration speed by more than 70%, reducing the average calibration time from 101.27 s to 30.99 s in the existing five datasets, but also maintains the calibration accuracy at the optimal level. The calibration error in the public dataset PlenCalCVPR2013DatasetA is only 0.0714 mm.
2022, 44(5): 1539-1547.
doi: 10.11999/JEIT211381
Abstract:
As an important carrier of intelligent manufacturing, industrial robot has great potential in large-scale and complex tasks. However, the problem of low positioning accuracy and difficulty to control hinders the further popularization of robots in high-precision tasks. In order to improve the accuracy of robot operation, a robot positioning error prediction and compensation method based on spatio-temporal convolution graph network is proposed in this work. Firstly, through the design of graph relation coding module and spatio-temporal feature decoding module, the prediction model of the robot position and orientation error based on graph convolution network is constructed; Then, to solve the problem of low efficiency caused by too many times of robotic inverse kinematics solution in traditional iteration compensation methods, the problem of compensation for positioning errors is transformed into optimization problem, and the genetic algorithm is used to compensate the position and attitude errors simultaneously; Finally, the training set is obtained by Latin hypercube sampling method to realize the training of robot positioning error prediction model, and the accuracy of positioning error prediction and the effect of compensation are verified by the experiments.
As an important carrier of intelligent manufacturing, industrial robot has great potential in large-scale and complex tasks. However, the problem of low positioning accuracy and difficulty to control hinders the further popularization of robots in high-precision tasks. In order to improve the accuracy of robot operation, a robot positioning error prediction and compensation method based on spatio-temporal convolution graph network is proposed in this work. Firstly, through the design of graph relation coding module and spatio-temporal feature decoding module, the prediction model of the robot position and orientation error based on graph convolution network is constructed; Then, to solve the problem of low efficiency caused by too many times of robotic inverse kinematics solution in traditional iteration compensation methods, the problem of compensation for positioning errors is transformed into optimization problem, and the genetic algorithm is used to compensate the position and attitude errors simultaneously; Finally, the training set is obtained by Latin hypercube sampling method to realize the training of robot positioning error prediction model, and the accuracy of positioning error prediction and the effect of compensation are verified by the experiments.
Defect Detection for Glass Seal Insulated Terminals Based on Sector Neighborhood Feature Engineering
2022, 44(5): 1548-1553.
doi: 10.11999/JEIT211346
Abstract:
In this paper, an appearance quality inspection method for glass seal insulated terminals is proposed based on feature engineering to replace the current manual inspection. First, the inspecting region in the glass seal insulated terminal image is divided into many sectors based on the shape prior of the terminal. Second, considering the characteristics of the image, four categories of sector features are designed, such as sector basic features, sector gray change rates, sector reflection features and sector direction statistical features. Then, they are input into a Gradient Boosting Decision Tree (GBDT) for rough classification. Next, to characterize excellently the sectors, a new image feature called Sector Neighborhood (SN) feature is designed by combining sector features of the nearest neighbor sectors and rough classification results for these sectors. Finally, the SN features and sector features are input into the GBDT for fine classification, which indicates final quality inspection. The experimental results indicate that the proposed method can achieve better inspection performance with reasonable inspection time compared to the existing inspection methods, which has 97.45% IoU and 0.987 F1.
In this paper, an appearance quality inspection method for glass seal insulated terminals is proposed based on feature engineering to replace the current manual inspection. First, the inspecting region in the glass seal insulated terminal image is divided into many sectors based on the shape prior of the terminal. Second, considering the characteristics of the image, four categories of sector features are designed, such as sector basic features, sector gray change rates, sector reflection features and sector direction statistical features. Then, they are input into a Gradient Boosting Decision Tree (GBDT) for rough classification. Next, to characterize excellently the sectors, a new image feature called Sector Neighborhood (SN) feature is designed by combining sector features of the nearest neighbor sectors and rough classification results for these sectors. Finally, the SN features and sector features are input into the GBDT for fine classification, which indicates final quality inspection. The experimental results indicate that the proposed method can achieve better inspection performance with reasonable inspection time compared to the existing inspection methods, which has 97.45% IoU and 0.987 F1.
2022, 44(5): 1554-1562.
doi: 10.11999/JEIT211378
Abstract:
In view of the previous flexible integrated scheduling algorithms that consider forward scheduling, so that it is necessary to consider the multiply predecessors’ constraints of the target operation, which makes it difficult to arrange rationally the relevant operations and affect the product completion time. A flexible integrated scheduling algorithm based on reverse order layer priority is proposed. Firstly, a reverse-order layer priority strategy is proposed, and each operation is allocated to the set of operations to be scheduled in the reverse-order layer; Secondly, a dynamic pseudo-long path strategy is proposed to determine the scheduling sequence of the centralized operation of each reverse-order layer to be scheduled; Then, the equipment selection strategy and the equipment preemption strategy are proposed to determine the processing equipment and processing time of the target operation. Finally, a scheduling plan conversion strategy based on the completion time flip is proposed, which converts the reverse order scheduling scheme into a positive order scheduling scheme. The example shows that compared with the existing mainstream algorithms, this algorithm can shorten the product completion time without increasing the complexity of the algorithm.
In view of the previous flexible integrated scheduling algorithms that consider forward scheduling, so that it is necessary to consider the multiply predecessors’ constraints of the target operation, which makes it difficult to arrange rationally the relevant operations and affect the product completion time. A flexible integrated scheduling algorithm based on reverse order layer priority is proposed. Firstly, a reverse-order layer priority strategy is proposed, and each operation is allocated to the set of operations to be scheduled in the reverse-order layer; Secondly, a dynamic pseudo-long path strategy is proposed to determine the scheduling sequence of the centralized operation of each reverse-order layer to be scheduled; Then, the equipment selection strategy and the equipment preemption strategy are proposed to determine the processing equipment and processing time of the target operation. Finally, a scheduling plan conversion strategy based on the completion time flip is proposed, which converts the reverse order scheduling scheme into a positive order scheduling scheme. The example shows that compared with the existing mainstream algorithms, this algorithm can shorten the product completion time without increasing the complexity of the algorithm.
2022, 44(5): 1563-1571.
doi: 10.11999/JEIT211358
Abstract:
In order to solve the problem that there is no real reference of the fabric image in the flexo first item inspection, a coarse-to-fine inspection method of flexo first item is proposed based on electronic samples, which is mainly divided into three stages: coarse matching, fine matching and defect detection. First, since different thickness of the characters, large differences in gray characteristics, and high repetition of flexo content inherently exist in the electronic sample and the flexo first item, the SuperGlue with SuperPoint are employed for rough matching. Then, a Normalized Cross-Correlation(NCC)-based method is used to fine-tune the electronic sample characters for fine matching, which can deal with local offsets of flexo content are caused by plate expansion and bending in flexo printing process. Finally, a constrained clustering method is proposed to transform the defect detection problem into the problem of minimizing the difference between electronic sample and flexo first item. Comparison experiments show that the proposed method can achieve better inspection performance for flexo first item, with the missed detection rate of 0, the false detection rate of 1.3%, the average Dice coefficient of 0.941, and the inspection time of 2.761 s/pcs. This promising inspection indicates that the proposed method can be well employed in real industries.
In order to solve the problem that there is no real reference of the fabric image in the flexo first item inspection, a coarse-to-fine inspection method of flexo first item is proposed based on electronic samples, which is mainly divided into three stages: coarse matching, fine matching and defect detection. First, since different thickness of the characters, large differences in gray characteristics, and high repetition of flexo content inherently exist in the electronic sample and the flexo first item, the SuperGlue with SuperPoint are employed for rough matching. Then, a Normalized Cross-Correlation(NCC)-based method is used to fine-tune the electronic sample characters for fine matching, which can deal with local offsets of flexo content are caused by plate expansion and bending in flexo printing process. Finally, a constrained clustering method is proposed to transform the defect detection problem into the problem of minimizing the difference between electronic sample and flexo first item. Comparison experiments show that the proposed method can achieve better inspection performance for flexo first item, with the missed detection rate of 0, the false detection rate of 1.3%, the average Dice coefficient of 0.941, and the inspection time of 2.761 s/pcs. This promising inspection indicates that the proposed method can be well employed in real industries.
2022, 44(5): 1572-1580.
doi: 10.11999/JEIT211455
Abstract:
For the general integrated scheduling problem of tree structured complex single product machining and assembling, a reverse order and greedy integrated scheduling algorithm is proposed by considering dynamic TUD (Time Urgency Degree) of the process sequences. The strategy of process sorting is put forward, and the TUD of process sequence is defined. The process tree is reversed using leaf alignment, according to the order from leaf to root, the scheduling order of leaf nodes in the same layer is determined layer by layer from large to small according to the dynamic TUD values of the process sequences to which the leaf nodes belong. The sorted leaf nodes are put into the queue in turn. Finally, the elements in the queue are reversed. A reverse order and greedy scheduling strategy is proposed. Each time, a single process is taken as a unit to conduct trial scheduling at the quasi-scheduling time point in the required equipment. Quasi-scheduling scheme set of the process is obtained, and the quasi-scheduling scheme with the minimum end time is selected, and if it is not unique, the scheme is selected to machine the process as early as possible. A case shows that the proposed algorithm optimizes the general integrated scheduling results and has high efficiency.
For the general integrated scheduling problem of tree structured complex single product machining and assembling, a reverse order and greedy integrated scheduling algorithm is proposed by considering dynamic TUD (Time Urgency Degree) of the process sequences. The strategy of process sorting is put forward, and the TUD of process sequence is defined. The process tree is reversed using leaf alignment, according to the order from leaf to root, the scheduling order of leaf nodes in the same layer is determined layer by layer from large to small according to the dynamic TUD values of the process sequences to which the leaf nodes belong. The sorted leaf nodes are put into the queue in turn. Finally, the elements in the queue are reversed. A reverse order and greedy scheduling strategy is proposed. Each time, a single process is taken as a unit to conduct trial scheduling at the quasi-scheduling time point in the required equipment. Quasi-scheduling scheme set of the process is obtained, and the quasi-scheduling scheme with the minimum end time is selected, and if it is not unique, the scheme is selected to machine the process as early as possible. A case shows that the proposed algorithm optimizes the general integrated scheduling results and has high efficiency.
2022, 44(5): 1581-1590.
doi: 10.11999/JEIT211414
Abstract:
Automatic monitoring of product quality has always been the core of intelligent development of injection molding industry. Practical factors like high cost of data collection, small sizes of sample and unbalanced sample categories require higher challenges for quality prediction of injection molded products. Therefore, a quality prediction model for injection molded products based on Broad Learning System (BLS) is proposed. Specifically, with the three-dimensional sizes of products as predicted targets, p-Norm is applied into the general BLS model to handle the problems of small samples and unbalanced data. The dataset from task two of the fourth industrial big data innovation competition is adopted. 192 parameter features are collected, among which 17 basic features, 4 derived features and 2 injection machine adjusting parameters are extracted as the input of the model via correlation analysis. The comparative experiments are then carried out between the proposed method and methods like Support Vector Machines (SVM), K-Nearest Neighbor (KNN), MultiLayer Perceptron (MLP) and BLS, with a respective sample size of 8300 data in the training and testing sets. Experimental results show that pN-BLS has the most accurate and fast effect of prediction. In practical defect detection applications, pN-BLS can predict abnormal data more accurately and has higher robustness.
Automatic monitoring of product quality has always been the core of intelligent development of injection molding industry. Practical factors like high cost of data collection, small sizes of sample and unbalanced sample categories require higher challenges for quality prediction of injection molded products. Therefore, a quality prediction model for injection molded products based on Broad Learning System (BLS) is proposed. Specifically, with the three-dimensional sizes of products as predicted targets, p-Norm is applied into the general BLS model to handle the problems of small samples and unbalanced data. The dataset from task two of the fourth industrial big data innovation competition is adopted. 192 parameter features are collected, among which 17 basic features, 4 derived features and 2 injection machine adjusting parameters are extracted as the input of the model via correlation analysis. The comparative experiments are then carried out between the proposed method and methods like Support Vector Machines (SVM), K-Nearest Neighbor (KNN), MultiLayer Perceptron (MLP) and BLS, with a respective sample size of 8300 data in the training and testing sets. Experimental results show that pN-BLS has the most accurate and fast effect of prediction. In practical defect detection applications, pN-BLS can predict abnormal data more accurately and has higher robustness.
2022, 44(5): 1591-1601.
doi: 10.11999/JEIT211000
Abstract:
Focusing on the robot grasping problem of point cloud targets in complex scenes with weak texture and scattered placement, a 6D pose estimation deep learning network is proposed. First, the complex scenes of the physical environment are simulated, where point cloud targets are randomly placed in multiple poses to generate a dataset with real labels; Further, a 6D pose estimation deep learning network model is designed, and a Multiscale Point Cloud Segmentation Net (MPCS-Net) is proposed to segment point cloud instances directly on the complete geometric point cloud, solving the dependence on RGB information and point cloud segmentation pre-processing. Then, the Multilayer Feature Pose Estimation Net (MFPE-Net) is proposed, which addresses effectively the pose estimation problem of symmetrical objects. Finally, the experimental results and analysis confirm that, compared with the traditional point cloud registration methods and the existing deep learning pose estimation methods of the segmented point cloud, the proposed method achieves higher accuracy and more stable performance. The preferable robustness in estimating the pose of symmetrical objects also proves its efficacy.
Focusing on the robot grasping problem of point cloud targets in complex scenes with weak texture and scattered placement, a 6D pose estimation deep learning network is proposed. First, the complex scenes of the physical environment are simulated, where point cloud targets are randomly placed in multiple poses to generate a dataset with real labels; Further, a 6D pose estimation deep learning network model is designed, and a Multiscale Point Cloud Segmentation Net (MPCS-Net) is proposed to segment point cloud instances directly on the complete geometric point cloud, solving the dependence on RGB information and point cloud segmentation pre-processing. Then, the Multilayer Feature Pose Estimation Net (MFPE-Net) is proposed, which addresses effectively the pose estimation problem of symmetrical objects. Finally, the experimental results and analysis confirm that, compared with the traditional point cloud registration methods and the existing deep learning pose estimation methods of the segmented point cloud, the proposed method achieves higher accuracy and more stable performance. The preferable robustness in estimating the pose of symmetrical objects also proves its efficacy.
2022, 44(5): 1602-1609.
doi: 10.11999/JEIT211434
Abstract:
Defect detection is an important part of intelligent manufacturing system. When traditional machine learning algorithms are used for defect classification, data noise interference is usually encountered, which reduces the algorithm’s prediction accuracy for defect classification. Although powerful algorithms such as Robust Linear Discriminant Analysis (RLDA) have been proposed in recent years to solve classification problems with data disturbed by sparse noise, there are still some drawbacks that limit its application performance. In this paper, a new Optimal Mean-Robust Linear Classification Analyis (OMRLSA) based on linear discriminant analysis is proposed. Different from the previous classification methods dealing with noisy data, ignoring the influence of the Laplace distribution characteristic of sparse noise on the data mean, the optimal mean robust linear classification model proposed in this paper will automatically update the optimal mean of the data. This ensures that the statistical characteristics of the data will not be disturbed by noise. Furthermore, a weighted kernel norm minimization method with joint L2,1 norm minimization and rank compression on regularization and error measurement is introduced for the first time in a robust classification model in the subsequent loss function. Thereby the robustness of the algorithm is improved. Experimental results on standard dataset with different ratio corruption illustrate the superiority of the proposed method.
Defect detection is an important part of intelligent manufacturing system. When traditional machine learning algorithms are used for defect classification, data noise interference is usually encountered, which reduces the algorithm’s prediction accuracy for defect classification. Although powerful algorithms such as Robust Linear Discriminant Analysis (RLDA) have been proposed in recent years to solve classification problems with data disturbed by sparse noise, there are still some drawbacks that limit its application performance. In this paper, a new Optimal Mean-Robust Linear Classification Analyis (OMRLSA) based on linear discriminant analysis is proposed. Different from the previous classification methods dealing with noisy data, ignoring the influence of the Laplace distribution characteristic of sparse noise on the data mean, the optimal mean robust linear classification model proposed in this paper will automatically update the optimal mean of the data. This ensures that the statistical characteristics of the data will not be disturbed by noise. Furthermore, a weighted kernel norm minimization method with joint L2,1 norm minimization and rank compression on regularization and error measurement is introduced for the first time in a robust classification model in the subsequent loss function. Thereby the robustness of the algorithm is improved. Experimental results on standard dataset with different ratio corruption illustrate the superiority of the proposed method.
2022, 44(5): 1610-1616.
doi: 10.11999/JEIT210857
Abstract:
For Hammerstein nonlinear time-varying systems running repeatedly on finite intervals, a weighted iterative learning algorithm is proposed to estimate the time-varying parameters involved in the system dynamics. The nonlinear input part of the Hammerstein system is tackled based on polynomial expansion, and the iterative learning least square algorithm is given for the time-varying parameter identification. In order to prevent data saturation, an iterative learning least squares algorithm with forgetting factor is proposed for reducing the system tracking error and improving the identification accuracy; A weighted iterative learning least squares algorithm is further presented by introducing the weight matrix. The derivations of the three algorithms are given in detail. The simulation results demonstrate the effectiveness of the proposed learning algorithms, and in comparison with iterative learning least squares algorithm, the modified one sreach high identification accuracy and need fewer iterations.
For Hammerstein nonlinear time-varying systems running repeatedly on finite intervals, a weighted iterative learning algorithm is proposed to estimate the time-varying parameters involved in the system dynamics. The nonlinear input part of the Hammerstein system is tackled based on polynomial expansion, and the iterative learning least square algorithm is given for the time-varying parameter identification. In order to prevent data saturation, an iterative learning least squares algorithm with forgetting factor is proposed for reducing the system tracking error and improving the identification accuracy; A weighted iterative learning least squares algorithm is further presented by introducing the weight matrix. The derivations of the three algorithms are given in detail. The simulation results demonstrate the effectiveness of the proposed learning algorithms, and in comparison with iterative learning least squares algorithm, the modified one sreach high identification accuracy and need fewer iterations.
2022, 44(5): 1617-1624.
doi: 10.11999/JEIT211246
Abstract:
In the process of reflow soldering, void defects inevitably emerge inside solder joints of chip resistors, which will influence reliability of the device. In this paper, an adaptive inspection method for void defects inside solder joints of chip resistors is proposed by combining a Local Pre-Fitted (LPF) active contour model and circular convolutions with adaptive kernels. First, since the image of chip resistor has two distinct regions, dark and bright regions are adaptively separated from the image after solving the optimization problem with the largest difference between the average gray level values of the two regions. Then, considering low contrast between voids and the image background, sparse distribution and large areas of voids in the dark region, LPF active contour model is used to inspect voids. As for the obvious difference between voids and the image background, dense distribution and small areas of voids in the bright region, circular convolutions with adaptive kernels are proposed to inspect voids. Finally, false detection can be eliminated by the shape factor and an average gray strategy to realize accurate void inspection. Experimental results show that the proposed method is superior to other inspection methods with an average Dice coefficient of 0.8846.
In the process of reflow soldering, void defects inevitably emerge inside solder joints of chip resistors, which will influence reliability of the device. In this paper, an adaptive inspection method for void defects inside solder joints of chip resistors is proposed by combining a Local Pre-Fitted (LPF) active contour model and circular convolutions with adaptive kernels. First, since the image of chip resistor has two distinct regions, dark and bright regions are adaptively separated from the image after solving the optimization problem with the largest difference between the average gray level values of the two regions. Then, considering low contrast between voids and the image background, sparse distribution and large areas of voids in the dark region, LPF active contour model is used to inspect voids. As for the obvious difference between voids and the image background, dense distribution and small areas of voids in the bright region, circular convolutions with adaptive kernels are proposed to inspect voids. Finally, false detection can be eliminated by the shape factor and an average gray strategy to realize accurate void inspection. Experimental results show that the proposed method is superior to other inspection methods with an average Dice coefficient of 0.8846.
2022, 44(5): 1625-1635.
doi: 10.11999/JEIT211366
Abstract:
In order to solve the problem that scheduling gaps between processes on equipment with many processes will have important influence on scheduling results in the integrated scheduling research of multi-variety and small batch complex products, a resource cooperative integrated scheduling algorithm considering the weight of equipment with many processes is proposed. In the integrated scheduling, the definition of multi-process equipment and process weight value is proposed for the first time, and then the scheduling strategy based on weight value is proposed to improve the tightness of longitudinal continuous processing. Finally, the optimal scheduling time adjustment strategy is proposed to improve the intensity of horizontal parallel optimization. Experimental results show that the algorithm has better performance in improving the overall utilization rate of the integrated scheduling equipment and reducing the time cost of complex products.
In order to solve the problem that scheduling gaps between processes on equipment with many processes will have important influence on scheduling results in the integrated scheduling research of multi-variety and small batch complex products, a resource cooperative integrated scheduling algorithm considering the weight of equipment with many processes is proposed. In the integrated scheduling, the definition of multi-process equipment and process weight value is proposed for the first time, and then the scheduling strategy based on weight value is proposed to improve the tightness of longitudinal continuous processing. Finally, the optimal scheduling time adjustment strategy is proposed to improve the intensity of horizontal parallel optimization. Experimental results show that the algorithm has better performance in improving the overall utilization rate of the integrated scheduling equipment and reducing the time cost of complex products.
2022, 44(5): 1636-1642.
doi: 10.11999/JEIT210870
Abstract:
The slight aesthetic defects of polarizing films can hardly image and are difficult to detect. A novel method of detecting the slight defects based on polarization imaging is proposed in this paper. The mechanism of contrast enhancement is described qualitatively through the measurement results of defect polarization index. The image contrast of the defect is greatly improved by making use of the significant difference of polarization state of the transmitted light between the defect and the normal region, so as to simplify the following image processing algorithm and improve both the detection speed and accuracy. The experimental results show that the average recognition rate of polarizer aesthetic defects is 97.3%, and the average detection time of a single defect sample is about 0.22 s, thus it meets basically the requirements of industrial application.
The slight aesthetic defects of polarizing films can hardly image and are difficult to detect. A novel method of detecting the slight defects based on polarization imaging is proposed in this paper. The mechanism of contrast enhancement is described qualitatively through the measurement results of defect polarization index. The image contrast of the defect is greatly improved by making use of the significant difference of polarization state of the transmitted light between the defect and the normal region, so as to simplify the following image processing algorithm and improve both the detection speed and accuracy. The experimental results show that the average recognition rate of polarizer aesthetic defects is 97.3%, and the average detection time of a single defect sample is about 0.22 s, thus it meets basically the requirements of industrial application.
2022, 44(5): 1643-1652.
doi: 10.11999/JEIT210784
Abstract:
A medical image segmentation network with boundary enhancement, named as the AS-UNet (Add-and-Subtract UNet), is proposed to solve the problems of traditional segmentation networks for medical images, such as unclear boundary segmentation and large missing value. The mask boundary image is obtained by using the mask boundary image extraction algorithm, and the Boundary Attention Block (BAB) with multi-scale feature maps is introduced into the last three layers of the UNet expansion path. Moreover, the combinatorial loss function is proposed to improve the segmentation accuracy. In testing, the BAB can be abandoned to reduce testing parameters. Comparisons with other segmentation methods on three different types of medical image segmentation datasets, Glas, DRIVE and ISIC2018 are provided, indicating that the segmentation performance of the AS-UNet is better.
A medical image segmentation network with boundary enhancement, named as the AS-UNet (Add-and-Subtract UNet), is proposed to solve the problems of traditional segmentation networks for medical images, such as unclear boundary segmentation and large missing value. The mask boundary image is obtained by using the mask boundary image extraction algorithm, and the Boundary Attention Block (BAB) with multi-scale feature maps is introduced into the last three layers of the UNet expansion path. Moreover, the combinatorial loss function is proposed to improve the segmentation accuracy. In testing, the BAB can be abandoned to reduce testing parameters. Comparisons with other segmentation methods on three different types of medical image segmentation datasets, Glas, DRIVE and ISIC2018 are provided, indicating that the segmentation performance of the AS-UNet is better.
2022, 44(5): 1653-1663.
doi: 10.11999/JEIT211394
Abstract:
There are restrictions on the use time of equipment in self-owned processing enterprises. This limitation makes the product impossible to complete within the due date of the integrated scheduling problem. To solve this problem, Integrated scheduling algorithm for two corporate synergies with equipment time constraints is proposed. In order to ensure the self-owned processing enterprises can obtain more profits, it is necessary to assign as many processing tasks as possible to the self-owned processing enterprises for processing. Therefore, the processing task needs to be effectively decomposed. First, the processing tree is traversed in reverse direction. The upper limit of equipment use time in self-owned processing enterprise is taken as the threshold value. The processing task assignment strategy is designed to split the processing tree and generate the processing tree of the self-owned processing enterprise and the rest part is the processing tree of the cooperative processing enterprise. The collaborative selection strategy is designed. Under the premise of considering the transportation problem and meeting the delivery time, the cooperative processing enterprise is selected as the enterprise that makes the most profit of the self-owned processing enterprises. The example shows that the algorithm can better solve the problem of integrated scheduling with due date and profit, which is limited by the use time of equipment.
There are restrictions on the use time of equipment in self-owned processing enterprises. This limitation makes the product impossible to complete within the due date of the integrated scheduling problem. To solve this problem, Integrated scheduling algorithm for two corporate synergies with equipment time constraints is proposed. In order to ensure the self-owned processing enterprises can obtain more profits, it is necessary to assign as many processing tasks as possible to the self-owned processing enterprises for processing. Therefore, the processing task needs to be effectively decomposed. First, the processing tree is traversed in reverse direction. The upper limit of equipment use time in self-owned processing enterprise is taken as the threshold value. The processing task assignment strategy is designed to split the processing tree and generate the processing tree of the self-owned processing enterprise and the rest part is the processing tree of the cooperative processing enterprise. The collaborative selection strategy is designed. Under the premise of considering the transportation problem and meeting the delivery time, the cooperative processing enterprise is selected as the enterprise that makes the most profit of the self-owned processing enterprises. The example shows that the algorithm can better solve the problem of integrated scheduling with due date and profit, which is limited by the use time of equipment.
2022, 44(5): 1664-1673.
doi: 10.11999/JEIT211419
Abstract:
In the process of injection molding, the injection speed control is one of the important aspects, and it is important to achieve fast and reliable optimal control of the injection speed for the efficient production of injection molded products. In this paper, an efficient hybrid intelligent optimal control method based on the combination of control parameterization and particle swarm optimization is proposed, and an open-loop optimal controller and a state feedback optimal controller are designed and implemented, respectively. The controller design problem is transformed into a sequence of optimal parameter selection problems to achieve an efficient solution for the desired injection speed tracking control in a given time. Finally, the feasibility and effectiveness of the proposed hybrid optimal control algorithm for solving the dynamic optimization problem of the injection speed in the injection molding process are verified by experimental simulation results.
In the process of injection molding, the injection speed control is one of the important aspects, and it is important to achieve fast and reliable optimal control of the injection speed for the efficient production of injection molded products. In this paper, an efficient hybrid intelligent optimal control method based on the combination of control parameterization and particle swarm optimization is proposed, and an open-loop optimal controller and a state feedback optimal controller are designed and implemented, respectively. The controller design problem is transformed into a sequence of optimal parameter selection problems to achieve an efficient solution for the desired injection speed tracking control in a given time. Finally, the feasibility and effectiveness of the proposed hybrid optimal control algorithm for solving the dynamic optimization problem of the injection speed in the injection molding process are verified by experimental simulation results.
2022, 44(5): 1674-1682.
doi: 10.11999/JEIT210298
Abstract:
Although deep learning-based methods show their superiority in the field of single image deblurring, it is difficult to be applied to practice for requiring more computing resources and memory consumption as network deepens. In this work, a lightweight and fast progressive residual network for image deburring is proposed. The network takes shallow residual network as basic model to make full use of the local feature information and strengthen the information flow during back propagation. By reusing the residual network recursively in subsequent several stages and sharing parameters, the network model can be greatly simplified and the parameters can be reduced. To improve the reconstruction performance of the network, the feature recalibration module is applied to feature fusion. The channel attention mechanism is applied to integrate input image and output feature map of each residual network, and then the spatial information of feature map is selected adaptively to achieve better feature reconstruction. Experimental results show that the proposed model has fast running speed with a small number of parameters, which is much better than the existing algorithms, and can produce quite promising results for the removal of spatial-invariant blurring.
Although deep learning-based methods show their superiority in the field of single image deblurring, it is difficult to be applied to practice for requiring more computing resources and memory consumption as network deepens. In this work, a lightweight and fast progressive residual network for image deburring is proposed. The network takes shallow residual network as basic model to make full use of the local feature information and strengthen the information flow during back propagation. By reusing the residual network recursively in subsequent several stages and sharing parameters, the network model can be greatly simplified and the parameters can be reduced. To improve the reconstruction performance of the network, the feature recalibration module is applied to feature fusion. The channel attention mechanism is applied to integrate input image and output feature map of each residual network, and then the spatial information of feature map is selected adaptively to achieve better feature reconstruction. Experimental results show that the proposed model has fast running speed with a small number of parameters, which is much better than the existing algorithms, and can produce quite promising results for the removal of spatial-invariant blurring.
2022, 44(5): 1683-1693.
doi: 10.11999/JEIT210247
Abstract:
Liver tumor segmentation from abdominal CT image is an important prerequisite for liver disease diagnosis, surgical planning, and radiation therapy. However, the segmentation remains a challenging problem since the tumors in CT images generally have heterogeneous intensities, complicated textures, and ambiguous boundaries. To address this, an automatic, accurate, and robust segmentation method is proposed based on cascaded Dense-Unet and graph cuts. Firstly, the cascaded Dense-UNet is used to obtain liver tumor initial segmentation results as well as the tumor Regions Of Interest (ROIs). Then, an intensity model and a probability model are established respectively by utilizing pixel-wise and patch-wise features in order to distinguish between tumor and non-tumor, and these models are further integrated into the graph cuts energy function to segment the tumor from ROIs accurately. Finally, experiments are carried out on LiTS and 3Dircadb datasets, which are respectively used as training set and testing set, and this method is compared with many other existing automatic segmentation methods. Results demonstrate that the proposed method can segment liver tumors in CT images with different intensity, texture, shape and size more effectively and can extract the tumor boundaries more accurately than other methods, especially for the tumors with low contrasts and ambiguous boundaries.
Liver tumor segmentation from abdominal CT image is an important prerequisite for liver disease diagnosis, surgical planning, and radiation therapy. However, the segmentation remains a challenging problem since the tumors in CT images generally have heterogeneous intensities, complicated textures, and ambiguous boundaries. To address this, an automatic, accurate, and robust segmentation method is proposed based on cascaded Dense-Unet and graph cuts. Firstly, the cascaded Dense-UNet is used to obtain liver tumor initial segmentation results as well as the tumor Regions Of Interest (ROIs). Then, an intensity model and a probability model are established respectively by utilizing pixel-wise and patch-wise features in order to distinguish between tumor and non-tumor, and these models are further integrated into the graph cuts energy function to segment the tumor from ROIs accurately. Finally, experiments are carried out on LiTS and 3Dircadb datasets, which are respectively used as training set and testing set, and this method is compared with many other existing automatic segmentation methods. Results demonstrate that the proposed method can segment liver tumors in CT images with different intensity, texture, shape and size more effectively and can extract the tumor boundaries more accurately than other methods, especially for the tumors with low contrasts and ambiguous boundaries.
2022, 44(5): 1694-1703.
doi: 10.11999/JEIT210389
Abstract:
Change detection in high resolution remote sensing images is the key to understanding of land surface changes. Change detection of remote sensing images is an important branch of remote sensing image processing. Many existing change detection methods based on deep learning have achieved good results, but it is not easy to obtain the structural details of high resolution remote sensing images, and the accuracy of the detection needs to be improved. Therefore, a network framework which combines Edge change information and channel Attention Network module (EANet) is proposed. EANet is divided into three modules: Edge structure change information detection, depth feature extraction and change area discrimination. Firstly, in order to get the edge change information of the two-phase images, the edge of the two-phase images is detected to get the edge images, and the edge images is subtracted to get the edge difference images. Secondly, in consideration of the fine image details and complex texture features of high resolution remote sensing images, in order to extract fully the depth features of a single image, a model with three branches based on VGG-16 network is constructed to extract the depth features of bitemporal images and edge difference images respectively. Finally, in order to improve the accuracy of the detection, the channel attention mechanism is embedded into the model to focus on the channel features with large amount of information to identify better the changed regions. The experimental results show that the proposed algorithm is superior to some existing methods in terms of visual interpretation and accuracy measurement.
Change detection in high resolution remote sensing images is the key to understanding of land surface changes. Change detection of remote sensing images is an important branch of remote sensing image processing. Many existing change detection methods based on deep learning have achieved good results, but it is not easy to obtain the structural details of high resolution remote sensing images, and the accuracy of the detection needs to be improved. Therefore, a network framework which combines Edge change information and channel Attention Network module (EANet) is proposed. EANet is divided into three modules: Edge structure change information detection, depth feature extraction and change area discrimination. Firstly, in order to get the edge change information of the two-phase images, the edge of the two-phase images is detected to get the edge images, and the edge images is subtracted to get the edge difference images. Secondly, in consideration of the fine image details and complex texture features of high resolution remote sensing images, in order to extract fully the depth features of a single image, a model with three branches based on VGG-16 network is constructed to extract the depth features of bitemporal images and edge difference images respectively. Finally, in order to improve the accuracy of the detection, the channel attention mechanism is embedded into the model to focus on the channel features with large amount of information to identify better the changed regions. The experimental results show that the proposed algorithm is superior to some existing methods in terms of visual interpretation and accuracy measurement.
2022, 44(5): 1704-1713.
doi: 10.11999/JEIT210232
Abstract:
When the standard Nonnegative Matrix Factorization (NMF) is applied to hyperspectral unmixing, it is easy to be interfered by noise and outliers, and the unmixing effect is poor. In order to improve the factorized performance, the L21 norm is introduced into the standard NMF algorithm, and the model is improved to improve the robustness of the algorithm. Secondly, in order to improve the sparsity of the factorized abundance matrix, the double reweighted sparse constraint is introduced into the L21NMF model, so that one of the weights increases sparsity along the abundance vector corresponding to each pixel, and the other weight promotes the sparsity along the abundance vector corresponding to each endmember. Meanwhile, in order to utilize the global spatial distribution information of the pixels and observe the true distribution of materials in different images, the subspace structure regularization is introduced, and the L21 Nonnegative Matrix Factorization based on Subspace Structure Regularization (L21NMF-SSR) is proposed. The better performance and denoising ability of the proposed method are demonstrated by comparing with other classical methods on both synthetic and real datasets.
When the standard Nonnegative Matrix Factorization (NMF) is applied to hyperspectral unmixing, it is easy to be interfered by noise and outliers, and the unmixing effect is poor. In order to improve the factorized performance, the L21 norm is introduced into the standard NMF algorithm, and the model is improved to improve the robustness of the algorithm. Secondly, in order to improve the sparsity of the factorized abundance matrix, the double reweighted sparse constraint is introduced into the L21NMF model, so that one of the weights increases sparsity along the abundance vector corresponding to each pixel, and the other weight promotes the sparsity along the abundance vector corresponding to each endmember. Meanwhile, in order to utilize the global spatial distribution information of the pixels and observe the true distribution of materials in different images, the subspace structure regularization is introduced, and the L21 Nonnegative Matrix Factorization based on Subspace Structure Regularization (L21NMF-SSR) is proposed. The better performance and denoising ability of the proposed method are demonstrated by comparing with other classical methods on both synthetic and real datasets.
2022, 44(5): 1714-1722.
doi: 10.11999/JEIT210296
Abstract:
The application of Siamese network to visual object tracking has greatly improved the performance of the tracker recently, which can take both accuracy and speed into account. However, the accuracy of Siamese network tracker is limited to a great extent. In order to solve the above problems, a key information feature perception module based on channel attention mechanism to enhance the discrimination ability of the network model is proposed, which make the network focus on the convolution feature changes of the target; On this basis, an online adaptive masking strategy is proposed, which adaptively masks the subsequent frames according to the output state of the cross-correlation layer learned online, so as to highlight the foreground target. Experiments on OTB100 and GOT-10k datasets show that without affecting the real-time performance, the proposed tracker has a significant improvement in accuracy compared with the benchmark, and has a robust tracking effect in complex scenes such as occlusion, scale change and background clutter.
The application of Siamese network to visual object tracking has greatly improved the performance of the tracker recently, which can take both accuracy and speed into account. However, the accuracy of Siamese network tracker is limited to a great extent. In order to solve the above problems, a key information feature perception module based on channel attention mechanism to enhance the discrimination ability of the network model is proposed, which make the network focus on the convolution feature changes of the target; On this basis, an online adaptive masking strategy is proposed, which adaptively masks the subsequent frames according to the output state of the cross-correlation layer learned online, so as to highlight the foreground target. Experiments on OTB100 and GOT-10k datasets show that without affecting the real-time performance, the proposed tracker has a significant improvement in accuracy compared with the benchmark, and has a robust tracking effect in complex scenes such as occlusion, scale change and background clutter.
2022, 44(5): 1723-1733.
doi: 10.11999/JEIT210230
Abstract:
At present, the study of knowledge graph focuses mainly on information retrieval, natural language understanding and other fields. Integrating knowledge graph with recommendation system has been concerned by scholars in the field. In order to mine the rich information ignored in knowledge graph, the knowledge graph is extended to multimodal and a recommendation model that incorporates Knowledge Graph with Image (KG-I) features is proposed. Different from other recommendation algorithms, visual embedding, knowledge embedding and structure embedding are combined to capture implicit feedback between user-items. The Deep Walk is used to capture the spatial structure and the ideal of RippleNet to retain the semantic features of knowledge graph, and the effect of images on preference is considered to integrate information. Compared with other models on the real data set, the influence of various features is studied, and the performance of different sparsity data is analyzed. The results show that the personalized recommendation model based on knowledge graph and image features outperforms other algorithms and the data sparsity can be alleviated.
At present, the study of knowledge graph focuses mainly on information retrieval, natural language understanding and other fields. Integrating knowledge graph with recommendation system has been concerned by scholars in the field. In order to mine the rich information ignored in knowledge graph, the knowledge graph is extended to multimodal and a recommendation model that incorporates Knowledge Graph with Image (KG-I) features is proposed. Different from other recommendation algorithms, visual embedding, knowledge embedding and structure embedding are combined to capture implicit feedback between user-items. The Deep Walk is used to capture the spatial structure and the ideal of RippleNet to retain the semantic features of knowledge graph, and the effect of images on preference is considered to integrate information. Compared with other models on the real data set, the influence of various features is studied, and the performance of different sparsity data is analyzed. The results show that the personalized recommendation model based on knowledge graph and image features outperforms other algorithms and the data sparsity can be alleviated.
2022, 44(5): 1734-1742.
doi: 10.11999/JEIT210220
Abstract:
In order to solve the problem of low accuracy and low efficiency of car loudspeakers sound field correction, a correction method of adaptive asymmetric parallel graphic equalizer without overlapping frequency bands is proposed in this paper. In the case of the dynamic change of the sound field in the car, the proposed method takes into account the effective equalization range and adaptive gains, rather than the fixed equalization range and given artificially gains in the classical methods. Through the experimental analysis of the measured data, the average number of equalization filters used by the proposed method is about 20% less than the classical methods, whose fitting target gains is more accurate, and the spectrum curve is flatter after correction.
In order to solve the problem of low accuracy and low efficiency of car loudspeakers sound field correction, a correction method of adaptive asymmetric parallel graphic equalizer without overlapping frequency bands is proposed in this paper. In the case of the dynamic change of the sound field in the car, the proposed method takes into account the effective equalization range and adaptive gains, rather than the fixed equalization range and given artificially gains in the classical methods. Through the experimental analysis of the measured data, the average number of equalization filters used by the proposed method is about 20% less than the classical methods, whose fitting target gains is more accurate, and the spectrum curve is flatter after correction.
2022, 44(5): 1743-1753.
doi: 10.11999/JEIT210313
Abstract:
To address the problem that Simultaneous Location And Mapping (SLAM) closed-loop detection algorithms are easily disturbed by complex environmental factors, resulting in large localization errors and low closed-loop detection accuracy, a bionic SLAM algorithm based on the interest tendency mechanism is proposed, inspired by the spatial cognitive mechanism of mammals. The grid cells are modelled using the Lateral Anti-Hebbian Network (LAHN), which improves the accuracy of the algorithm by correcting the grid cells with irregular and complex environmental boundary information. The tendency of interest mechanism is used to score the extracted areas of significance, reduce the impact of redundant significant areas and improve the accuracy of the system’s closed-loop detection. A cognitive map is constructed by correlating the location information obtained from the location-aware model with a visual perception template. The results of the tests on the public dataset and the real environment show that the proposed algorithm has advantages in terms of accuracy, real time performance and adaptability to the environment.
To address the problem that Simultaneous Location And Mapping (SLAM) closed-loop detection algorithms are easily disturbed by complex environmental factors, resulting in large localization errors and low closed-loop detection accuracy, a bionic SLAM algorithm based on the interest tendency mechanism is proposed, inspired by the spatial cognitive mechanism of mammals. The grid cells are modelled using the Lateral Anti-Hebbian Network (LAHN), which improves the accuracy of the algorithm by correcting the grid cells with irregular and complex environmental boundary information. The tendency of interest mechanism is used to score the extracted areas of significance, reduce the impact of redundant significant areas and improve the accuracy of the system’s closed-loop detection. A cognitive map is constructed by correlating the location information obtained from the location-aware model with a visual perception template. The results of the tests on the public dataset and the real environment show that the proposed algorithm has advantages in terms of accuracy, real time performance and adaptability to the environment.
2022, 44(5): 1754-1762.
doi: 10.11999/JEIT210270
Abstract:
In order to enhance the security of image data transmission, a novel two-dimensional exponent - cosine chaotic map is proposed. In this system, a new chaotic map is constructed by introducing exponent and high-power nonlinear terms into one dimensional cosine chaotic system. The nonlinear term is introduced to perturb the iterative process of one-dimensional cosine chaotic system to obtain fuller chaotic phase orbits. Lyapunov exponential spectrum and system bifurcation diagram are used to verify the features of chaotic system. Based on the chaotic map, a novel image encryption algorithm is proposed. The encrypted data has good encryption security by following a “scrambling-diffusion-scrambling” strategy. Security analysis of encrypted image data also shows that the two-dimensional exponential - cosine chaotic map has strong robustness and encryption security.
In order to enhance the security of image data transmission, a novel two-dimensional exponent - cosine chaotic map is proposed. In this system, a new chaotic map is constructed by introducing exponent and high-power nonlinear terms into one dimensional cosine chaotic system. The nonlinear term is introduced to perturb the iterative process of one-dimensional cosine chaotic system to obtain fuller chaotic phase orbits. Lyapunov exponential spectrum and system bifurcation diagram are used to verify the features of chaotic system. Based on the chaotic map, a novel image encryption algorithm is proposed. The encrypted data has good encryption security by following a “scrambling-diffusion-scrambling” strategy. Security analysis of encrypted image data also shows that the two-dimensional exponential - cosine chaotic map has strong robustness and encryption security.
2022, 44(5): 1763-1772.
doi: 10.11999/JEIT210289
Abstract:
Incremental search services in search engines update the suggestion list for users by sending real-time requests. Focusing on the information leakage of encrypted search traffic, a side-channel analysis method on Chinese search is proposed. Leveraging the distinguishability of packet length increments and time intervals, a three-stage analysis model is constructed to identify user queries. Experimental results show that the performance in four commonly used Chinese search engines achieves the theoretical quantified value. The identification accuracy for the set containing 1.4×105 monitored queries reaches 76%. Finally, four mitigation methods are evaluated to demonstrate that side-channel analysis can be effectively defended by blocking the information leakage sources.
Incremental search services in search engines update the suggestion list for users by sending real-time requests. Focusing on the information leakage of encrypted search traffic, a side-channel analysis method on Chinese search is proposed. Leveraging the distinguishability of packet length increments and time intervals, a three-stage analysis model is constructed to identify user queries. Experimental results show that the performance in four commonly used Chinese search engines achieves the theoretical quantified value. The identification accuracy for the set containing 1.4×105 monitored queries reaches 76%. Finally, four mitigation methods are evaluated to demonstrate that side-channel analysis can be effectively defended by blocking the information leakage sources.
2022, 44(5): 1773-1781.
doi: 10.11999/JEIT210276
Abstract:
Gradient redundancy exists in the process of deep neural network gradient descent. When differential privacy mechanism is applied to resist member inference attack, excessive noise will be introduced. So, the gradient matrix is decomposed by Funk-SVD algorithm and noise is added to the low-dimensional eigen subspace matrix and residual matrix respectively. The redundant gradient noise is eliminated in the gradient reconstruction process. The decomposition matrix norm is recalculated and the smoothing sensitivity is combined to reduce the noise scale. At the same time, according to the correlation between input features and output features, more privacy budget is allocated to features with large correlation coefficients to improve the training accuracy. The noise scale is reduced by recalculating the decomposition matrix norm and the smoothing sensitivity. Moment accountant is used to calculate the cumulative privacy loss under multiple optimization strategies. The results show that Deep neural networks under differential privacy based on Funk-SVD (FSDP) can bridge the gap with the non-privacy model more effectively on MNIST and CIFAR-10.
Gradient redundancy exists in the process of deep neural network gradient descent. When differential privacy mechanism is applied to resist member inference attack, excessive noise will be introduced. So, the gradient matrix is decomposed by Funk-SVD algorithm and noise is added to the low-dimensional eigen subspace matrix and residual matrix respectively. The redundant gradient noise is eliminated in the gradient reconstruction process. The decomposition matrix norm is recalculated and the smoothing sensitivity is combined to reduce the noise scale. At the same time, according to the correlation between input features and output features, more privacy budget is allocated to features with large correlation coefficients to improve the training accuracy. The noise scale is reduced by recalculating the decomposition matrix norm and the smoothing sensitivity. Moment accountant is used to calculate the cumulative privacy loss under multiple optimization strategies. The results show that Deep neural networks under differential privacy based on Funk-SVD (FSDP) can bridge the gap with the non-privacy model more effectively on MNIST and CIFAR-10.
2022, 44(5): 1782-1791.
doi: 10.11999/JEIT210280
Abstract:
The existing image information hiding methods based on Deep Convolutional Neural Networks (DCNN) have the problems of poor image visual quality and low hiding capacity. Addressing such issues, an image hiding method based on a two-channel deep convolutional neural network is proposed. First, different from the previous hiding framework, the hiding method proposed in this paper includes one hiding network and two revealing networks with the same structure, and two full-size secret images can be effectively hidden and revealed at the same time is realized. Then, to improve the visual quality of the image, an improved pyramid pooling module and a preprocessing module are added to the hiding and revealing network. The test results on multiple data sets show that the proposed method has a significant improvement in visual quality compared with existing image information hiding methods. The Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity (SSIM) values are increased by 3.75 dB and 3.61 % respectively, a relative capacity of 2 and good generalization ability are achieved.
The existing image information hiding methods based on Deep Convolutional Neural Networks (DCNN) have the problems of poor image visual quality and low hiding capacity. Addressing such issues, an image hiding method based on a two-channel deep convolutional neural network is proposed. First, different from the previous hiding framework, the hiding method proposed in this paper includes one hiding network and two revealing networks with the same structure, and two full-size secret images can be effectively hidden and revealed at the same time is realized. Then, to improve the visual quality of the image, an improved pyramid pooling module and a preprocessing module are added to the hiding and revealing network. The test results on multiple data sets show that the proposed method has a significant improvement in visual quality compared with existing image information hiding methods. The Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity (SSIM) values are increased by 3.75 dB and 3.61 % respectively, a relative capacity of 2 and good generalization ability are achieved.
2022, 44(5): 1792-1799.
doi: 10.11999/JEIT210239
Abstract:
The Internet of vehicles application scenarios put forward higher requirements for wireless communication in terms of bandwidth, delay, and reliability, especially in the Vehicle to Vehicle (V2V) communication scenario. For the technical challenges of channel estimation caused by time/frequency domain selective fading (dual selection fading) and non-stationary characteristics in the V2V high-speed mobile scenario, this paper proposes a channel estimation method of BEM (Basis Expansion Model)-based UKF-RTSS (Unscented Kalman Filter-Rauch-Tung-Striebel Smoother). The BEM model is used to fit the time-varying channel, and the estimation of the channel parameters is converted into the estimation of the basis function coefficients; The unscented Kalman filter (UKF) algorithm is used to estimate jointly the channel impulse response and the time-varying time-domain autocorrelation coefficient at the data, tracking the fast time-varying channel response. In order to improve further the accuracy of channel estimation, RTSS is introduced to perform channel estimation and interpolation on the backward channel state information, and it forms a joint estimator with a "filtering and smoothing" structure with UKF. System simulation analysis shows that under different speed and time-varying conditions, the BEM-based UKF-RTSS channel estimation method has higher channel estimation accuracy and robustness than other classic methods, and is especially suitable for wireless communication in the Internet of vehicles scenarios.
The Internet of vehicles application scenarios put forward higher requirements for wireless communication in terms of bandwidth, delay, and reliability, especially in the Vehicle to Vehicle (V2V) communication scenario. For the technical challenges of channel estimation caused by time/frequency domain selective fading (dual selection fading) and non-stationary characteristics in the V2V high-speed mobile scenario, this paper proposes a channel estimation method of BEM (Basis Expansion Model)-based UKF-RTSS (Unscented Kalman Filter-Rauch-Tung-Striebel Smoother). The BEM model is used to fit the time-varying channel, and the estimation of the channel parameters is converted into the estimation of the basis function coefficients; The unscented Kalman filter (UKF) algorithm is used to estimate jointly the channel impulse response and the time-varying time-domain autocorrelation coefficient at the data, tracking the fast time-varying channel response. In order to improve further the accuracy of channel estimation, RTSS is introduced to perform channel estimation and interpolation on the backward channel state information, and it forms a joint estimator with a "filtering and smoothing" structure with UKF. System simulation analysis shows that under different speed and time-varying conditions, the BEM-based UKF-RTSS channel estimation method has higher channel estimation accuracy and robustness than other classic methods, and is especially suitable for wireless communication in the Internet of vehicles scenarios.
2022, 44(5): 1800-1808.
doi: 10.11999/JEIT210590
Abstract:
With the arrival of the Internet of Things (IoT) era, the problem of wireless network saturation has become more and more serious. In order to overcome this problem, the IEEE Standards Association (IEEE-SA) has formulated the latest standard for wireless local area networks—IEEE 802.11ax. In this standard, the Orthogonal Frequency Division Multiple Access (OFDMA) technology is utilized to divide wireless channel into several groups of tones, and the divided sub-channels are called Resource Units (RUs). In order to solve the channel resource scheduling problem of 802.11ax uplink in dense user environments, an RU scheduling algorithm based on reinforcement learning is proposed in this paper. The Actor-Critic algorithm is used to train the pointer network and solve the adaptive allocation problem of RU. Finally, RUs are allocated to each user reasonably with the guarantee of priority and fairness. The simulation results show that the scheduling algorithm is more effective than traditional scheduling methods in the IEEE 802.11ax uplink and has a strong generalization ability, which is suitable for the IoT scenario in dense user environments.
With the arrival of the Internet of Things (IoT) era, the problem of wireless network saturation has become more and more serious. In order to overcome this problem, the IEEE Standards Association (IEEE-SA) has formulated the latest standard for wireless local area networks—IEEE 802.11ax. In this standard, the Orthogonal Frequency Division Multiple Access (OFDMA) technology is utilized to divide wireless channel into several groups of tones, and the divided sub-channels are called Resource Units (RUs). In order to solve the channel resource scheduling problem of 802.11ax uplink in dense user environments, an RU scheduling algorithm based on reinforcement learning is proposed in this paper. The Actor-Critic algorithm is used to train the pointer network and solve the adaptive allocation problem of RU. Finally, RUs are allocated to each user reasonably with the guarantee of priority and fairness. The simulation results show that the scheduling algorithm is more effective than traditional scheduling methods in the IEEE 802.11ax uplink and has a strong generalization ability, which is suitable for the IoT scenario in dense user environments.
2022, 44(5): 1809-1818.
doi: 10.11999/JEIT210209
Abstract:
The security performance of Multiple-Input Single-Output (MISO) system with the aid of the Intelligent Reflecting Surface (IRS) is analyzed in this paper, where eavesdropping nodes are randomly distributed. The stochastic geometry theory is utilized to model the eavesdropping nodes as a homogeneous Poisson Point Process (PPP). With the transmit antenna selection strategy, the legitimate node selects the optimal link to transmit signal. And the phase shifts at the IRS are tuned to enhance the selected link quality. Then, considering the transmission secrecy outage probability as the performance metric, the closed expression of scheme is derived. Further, the impact of the parameters, such as the number of reflection units and transmitting antennas, on the outage probability is analyzed. Finally, the design strategy of parameters for maximizing the security performance is given. The simulation results verify the theoretical analysis and show that the proposed scheme can improve the security performance under low energy consumption.
The security performance of Multiple-Input Single-Output (MISO) system with the aid of the Intelligent Reflecting Surface (IRS) is analyzed in this paper, where eavesdropping nodes are randomly distributed. The stochastic geometry theory is utilized to model the eavesdropping nodes as a homogeneous Poisson Point Process (PPP). With the transmit antenna selection strategy, the legitimate node selects the optimal link to transmit signal. And the phase shifts at the IRS are tuned to enhance the selected link quality. Then, considering the transmission secrecy outage probability as the performance metric, the closed expression of scheme is derived. Further, the impact of the parameters, such as the number of reflection units and transmitting antennas, on the outage probability is analyzed. Finally, the design strategy of parameters for maximizing the security performance is given. The simulation results verify the theoretical analysis and show that the proposed scheme can improve the security performance under low energy consumption.
2022, 44(5): 1819-1827.
doi: 10.11999/JEIT210267
Abstract:
To investigate the in-body transmission mechanism and propose a comprehensive channel model at Human Body Communication (HBC) band, two types of human body model, that is, an anatomical numerical human model and a multi-layer heterogeneous geometric human model, are adopted to derive the path loss characteristics using electromagnetic simulations. The average path losses from the human heart transmitter to each receiver node on the body surface are first calculated by the Time-Domain Finite Integration method incorporated with the human model. A comprehensive logarithmic path loss model with a linear regulation term is proposed based on the surface wave propagation mechanism to describe the channel characteristics accurately. The shadow fading in decibel is found to follow Normal distribution. The simulation and experimental measurement results demonstrate that the proposed path loss model can well characterize the implant channel characteristics for 10～50 MHz HBC band signals. Using the anatomical numerical human model to perform the channel modeling and channel characteristics study in this frequency band can improve the accuracy and reliability of the constructed implant channel model. This work is the first time to construct the in-body to on-body path loss model for ultra-wideband 10～50 MHz signals at the HBC band.
To investigate the in-body transmission mechanism and propose a comprehensive channel model at Human Body Communication (HBC) band, two types of human body model, that is, an anatomical numerical human model and a multi-layer heterogeneous geometric human model, are adopted to derive the path loss characteristics using electromagnetic simulations. The average path losses from the human heart transmitter to each receiver node on the body surface are first calculated by the Time-Domain Finite Integration method incorporated with the human model. A comprehensive logarithmic path loss model with a linear regulation term is proposed based on the surface wave propagation mechanism to describe the channel characteristics accurately. The shadow fading in decibel is found to follow Normal distribution. The simulation and experimental measurement results demonstrate that the proposed path loss model can well characterize the implant channel characteristics for 10～50 MHz HBC band signals. Using the anatomical numerical human model to perform the channel modeling and channel characteristics study in this frequency band can improve the accuracy and reliability of the constructed implant channel model. This work is the first time to construct the in-body to on-body path loss model for ultra-wideband 10～50 MHz signals at the HBC band.
2022, 44(5): 1828-1838.
doi: 10.11999/JEIT210263
Abstract:
To improve further the information transmission rate and Bit Error Rate (BER) performance of existing multi-user chaos keying systems, an Orthogonal Multi-User Correlation Delay Differential Chaos Shift Keying (OMU-CD-DCSK) system is proposed in this paper. The system is based on DCSK combined with Correlation Delay Shift Keying (CDSK), which can transmit N bit multi-user information in each time slot using orthogonal Walsh code sequence, and then increase further the transmission rate by quadrature modulation technique. At the receiver, a moving average filter is used to reduce the noise variance and improve the BER performance, followed by correlation demodulation to recover the multi-user information bits. The theoretical BER of the system under multipath Rayleigh fading channel is derived, and verified by Monte Carlo simulation experiments. Furthermore, the integrated utility of the system is defined for evaluating the integrated performance of the chaotic system. Compare with other chaotic keying systems, the integrated performance of OMU-CD-DCSK has obvious advantages. Therefore, it is of great value in application.
To improve further the information transmission rate and Bit Error Rate (BER) performance of existing multi-user chaos keying systems, an Orthogonal Multi-User Correlation Delay Differential Chaos Shift Keying (OMU-CD-DCSK) system is proposed in this paper. The system is based on DCSK combined with Correlation Delay Shift Keying (CDSK), which can transmit N bit multi-user information in each time slot using orthogonal Walsh code sequence, and then increase further the transmission rate by quadrature modulation technique. At the receiver, a moving average filter is used to reduce the noise variance and improve the BER performance, followed by correlation demodulation to recover the multi-user information bits. The theoretical BER of the system under multipath Rayleigh fading channel is derived, and verified by Monte Carlo simulation experiments. Furthermore, the integrated utility of the system is defined for evaluating the integrated performance of the chaotic system. Compare with other chaotic keying systems, the integrated performance of OMU-CD-DCSK has obvious advantages. Therefore, it is of great value in application.
2022, 44(5): 1839-1846.
doi: 10.11999/JEIT210228
Abstract:
For a wireless powered MEC (Mobile Edge Computing) network, the execution delay as the time for data offloading and data execution is defined, and a multidimensional resource allocation scheme is proposed to minimize the execution delay of all nodes. Firstly, an execution delay minimization based multidimensional optimization problem is formulated by jointly optimizing the operation time of a power beacon, the portions of task bits for local computing and offloading, the computing frequency and the transmit power of per node, subject to the energy-causality constraint of nodes. As the formulated optimization problem includes couplings among optimization variables and the max-max function, it is non-convex and can not be solved by the existing convex tools. Therefore, a series of slack variables and auxiliary variables are introduced to simplify the optimization problem and decouple the coupled variables. Then after carefully inspecting the structure of the simplified problem, a dichotomy based iterative algorithm is proposed to obtain the optimal solution. Finally, computer simulations validate the correctness of the devised iterative algorithm and the advantages of the proposed resource allocation in terms of the execution delay.
For a wireless powered MEC (Mobile Edge Computing) network, the execution delay as the time for data offloading and data execution is defined, and a multidimensional resource allocation scheme is proposed to minimize the execution delay of all nodes. Firstly, an execution delay minimization based multidimensional optimization problem is formulated by jointly optimizing the operation time of a power beacon, the portions of task bits for local computing and offloading, the computing frequency and the transmit power of per node, subject to the energy-causality constraint of nodes. As the formulated optimization problem includes couplings among optimization variables and the max-max function, it is non-convex and can not be solved by the existing convex tools. Therefore, a series of slack variables and auxiliary variables are introduced to simplify the optimization problem and decouple the coupled variables. Then after carefully inspecting the structure of the simplified problem, a dichotomy based iterative algorithm is proposed to obtain the optimal solution. Finally, computer simulations validate the correctness of the devised iterative algorithm and the advantages of the proposed resource allocation in terms of the execution delay.
2022, 44(5): 1847-1855.
doi: 10.11999/JEIT210204
Abstract:
In beam space millimeter-wave massive Multi-Input Multi-Output (MIMO) system, the power leakage problem will lead to energy loss. To mitigate this problem, Minimum Phase Error based Beam Rotating (MPE-BR) precoding scheme is proposed. Firstly, the phase shifter-based beam selection network is adopted, the beam selection set is constructed such that each Radio Frequency (RF) chain selects multiple beams collect the leaked power in system. Then, the beam rotation combination scheme based on minimum phase is proposed. Maximum gain beam is taken as reference. The phases of the beam selection set are determined by minimum phase error criterion such that the channel gains of the selected beams are approximatively aligned in the same direction for maximizing the Signal-to-Noise Ratio (SNR) of each user. System performance is improved. Furthermore, the proposed precoding algorithm is theoretically analyzed. The expression of the upper bound of spectrum efficiency and energy efficiency are given. The correctness of the theoretical derivation is verified in experiment, and the performance of proposed method is close to the ideal case of no-leakage power. The proposed scheme obtains better spectrum efficiency and energy efficiency performance than the existing algorithms.
In beam space millimeter-wave massive Multi-Input Multi-Output (MIMO) system, the power leakage problem will lead to energy loss. To mitigate this problem, Minimum Phase Error based Beam Rotating (MPE-BR) precoding scheme is proposed. Firstly, the phase shifter-based beam selection network is adopted, the beam selection set is constructed such that each Radio Frequency (RF) chain selects multiple beams collect the leaked power in system. Then, the beam rotation combination scheme based on minimum phase is proposed. Maximum gain beam is taken as reference. The phases of the beam selection set are determined by minimum phase error criterion such that the channel gains of the selected beams are approximatively aligned in the same direction for maximizing the Signal-to-Noise Ratio (SNR) of each user. System performance is improved. Furthermore, the proposed precoding algorithm is theoretically analyzed. The expression of the upper bound of spectrum efficiency and energy efficiency are given. The correctness of the theoretical derivation is verified in experiment, and the performance of proposed method is close to the ideal case of no-leakage power. The proposed scheme obtains better spectrum efficiency and energy efficiency performance than the existing algorithms.
2022, 44(5): 1856-1864.
doi: 10.11999/JEIT210269
Abstract:
To improve the angle estimation performance of the MIMO (Multiple-Input-Multiple-Output) radar, the optimization of the transmit weighting matrix by setting the 2-norm error between the actual synthesized transmitting steering vector and the desired one as the objective function is studied in this paper. The maximization of the transmit power utilization can be enforced via imposing the uniform elemental power constraint on the formulation. Furthermore, a method based on the cyclic algorithm and improved Projection Descent and Retraction (PDR) is provided to settle the equivalent problem under the vectorization of the transmit weighting matrix. The closed solution can be achieved at each iteration, then the computational complexity is low of the proposed method. And the convergence can also be proved. The proposed method obtains the superior performance in angle estimation of MIMO radar based on focusing the transmit power into the desired spatial sector. Finally, simulation results are presented to verify the efficiency of the proposed method.
To improve the angle estimation performance of the MIMO (Multiple-Input-Multiple-Output) radar, the optimization of the transmit weighting matrix by setting the 2-norm error between the actual synthesized transmitting steering vector and the desired one as the objective function is studied in this paper. The maximization of the transmit power utilization can be enforced via imposing the uniform elemental power constraint on the formulation. Furthermore, a method based on the cyclic algorithm and improved Projection Descent and Retraction (PDR) is provided to settle the equivalent problem under the vectorization of the transmit weighting matrix. The closed solution can be achieved at each iteration, then the computational complexity is low of the proposed method. And the convergence can also be proved. The proposed method obtains the superior performance in angle estimation of MIMO radar based on focusing the transmit power into the desired spatial sector. Finally, simulation results are presented to verify the efficiency of the proposed method.
2022, 44(5): 1865-1878.
doi: 10.11999/JEIT210932
Abstract:
Deep learning model facing the non-independent and identically distributed data streams, the old knowledge will be covered by new knowledge, resulting in a significant performance degradation of model. Continuous Learning(CL) can acquire incremental available knowledge from non-independent and identically distributed data streams, continuously accumulate new knowledge without learning from scratch, and achieve human intelligence by imitating brain learning and memory mechanisms. In this paper, the brain-inspired continuous learning methods are reviewed. Firstly, the history of continuous learning is reviewed. Secondly, from the perspective of brain continuous learning mechanism, the research methods of continuous learning are divided into general methods and brain-inspired methods .The current research status of replay, regularization and sparsity, which are commonly used as the methods of continuous learning, are summarized, and their difficulties are analyzed under the existing technical conditions. To this end, four types of brain-inspired methods: synaptic, dual system, sleep and modularization, which are closer to the ability of brain continuous learning, are meticulously analyzed and compared . Finally, the application status of brain-inspire continuous learning are summarized, and the challenges and development of brain-inspire continuous learning under the existing technical conditions are discussed.
Deep learning model facing the non-independent and identically distributed data streams, the old knowledge will be covered by new knowledge, resulting in a significant performance degradation of model. Continuous Learning(CL) can acquire incremental available knowledge from non-independent and identically distributed data streams, continuously accumulate new knowledge without learning from scratch, and achieve human intelligence by imitating brain learning and memory mechanisms. In this paper, the brain-inspired continuous learning methods are reviewed. Firstly, the history of continuous learning is reviewed. Secondly, from the perspective of brain continuous learning mechanism, the research methods of continuous learning are divided into general methods and brain-inspired methods .The current research status of replay, regularization and sparsity, which are commonly used as the methods of continuous learning, are summarized, and their difficulties are analyzed under the existing technical conditions. To this end, four types of brain-inspired methods: synaptic, dual system, sleep and modularization, which are closer to the ability of brain continuous learning, are meticulously analyzed and compared . Finally, the application status of brain-inspire continuous learning are summarized, and the challenges and development of brain-inspire continuous learning under the existing technical conditions are discussed.
