基于多尺度卷积神经网络的自适应熵加权决策融合船舶图像分类方法

任永梅; 杨杰; 郭志强; 曹辉

doi:10.11999/JEIT200102

基于多尺度卷积神经网络的自适应熵加权决策融合船舶图像分类方法

doi: 10.11999/JEIT200102

任永梅^{1, 2},
杨杰^1, ,,
郭志强¹,
曹辉¹

1.
武汉理工大学信息工程学院宽带无线通信与传感器网络湖北省重点实验室武汉 430070
2.
湖南工学院电气与信息工程学院衡阳 421002

基金项目: 国家自然科学基金 (51879211)，国家重点研发计划(2020YFB1710800)，湖南省教育厅科学研究项目(18C0900)

详细信息

作者简介:
任永梅：女，1988年生，博士生，研究方向为图像处理与模式识别

杨杰：女，1960年生，教授，研究方向为图像处理与模式识别

郭志强：男，1976年生，教授，研究方向为图像处理与模式识别

曹辉：男，1986年生，讲师，研究方向为数字信号的智能处理

通讯作者:
杨杰　jieyang@whut.edu.cn

中图分类号: TN911.73; TP391.4
计量
- 文章访问数: 1910
- HTML全文浏览量: 610
- PDF下载量: 183
- 被引次数: 0
出版历程
- 收稿日期: 2020-02-11
- 修回日期: 2020-10-28
- 网络出版日期: 2020-11-16
- 刊出日期: 2021-05-18

Self-adaptive Entropy Weighted Decision Fusion Method for Ship Image Classification Based on Multi-scale Convolutional Neural Network

Yongmei REN^{1, 2},
Jie YANG^{1
, ,},
Zhiqiang GUO¹,
Hui CAO¹

1.
Hubei Key Laboratory of Broadband Wireless Communication and Sensor Networks, School of Information Engineering, Wuhan University of Technology, Wuhan 430070, China
2.
School of Electrical and Information Engineering, Hunan Institute of Technology, Hengyang 421002, China

Funds: The National Natural Science Foundation of China (51879211), The National Key Research and Development Program of China (2020YFB1710800), The Scientific Research Project of the Hunan Provincial Education Department (18C0900)

摘要

摘要: 针对单一尺度卷积神经网络(CNN)对船舶图像分类的局限性，该文提出一种多尺度CNN自适应熵加权决策融合方法用于船舶图像分类。首先使用多尺度CNN提取不同尺寸的船舶图像的多尺度特征，并训练得到不同子网络的最优模型；接着利用测试集船舶图像在最优模型上测试，得到多尺度CNN的Softmax函数输出的概率值，并计算得到信息熵，进而实现对不同输入船舶图像赋予自适应的融合权重；最后对不同子网络的Softmax函数输出概率值进行自适应熵加权决策融合实现船舶图像的最终分类。在VAIS数据集和自建数据集上分别进行了实验，提出的方法的分类准确率分别达到了95.07%和97.50%，实验结果表明，与单一尺度CNN分类方法以及其他较新方法相比，所提方法具有更优的分类性能。
- 图像处理 /
- 船舶图像分类 /
- 多尺度卷积神经网络 /
- 熵 /
- 决策融合
Abstract: Considering the limitation of single scale Convolutional Neural Network (CNN) for ship image classification, a self-adaptive entropy weighted decision fusion method for ship image classification based on multi-scale CNN is proposed. Firstly, the multi-scale CNN is used to extract the multi-scale features of ship image with different sizes, and the optimum models of different sub-networks are trained. Then, the ship images of test set are tested on the optimum models, and the probability value that is output by Softmax function of multi-scale CNN is obtained, which is used to calculate the information entropy so as to realize the adaptive weight assigned to different input ship images. Finally, self-adaptive entropy weighted decision fusion is carried out for the probability value that is output by Softmax function of different sub-networks to realize the final ship image classification. Experiments perform on VAIS (Visible And Infrared Spectrums) and self-built datasets respectively, and the proposed method achieves average accuracy of 95.07% and 97.50% on these datasets respectively. The experimental results show that the proposed method has better classification performance than those of the single scale CNN classification method and other state-of-the-art methods.
- Image processing /
- Ship image classification /
- Multi-scale Convolutional Neural Network (CNN) /
- Entropy /
- Decision fusion

HTML全文

图 1 多尺度卷积神经网络

下载: 全尺寸图片幻灯片

图 2 CNN1网络结构和参数

下载: 全尺寸图片幻灯片

图 3 CNN2网络结构和参数

下载: 全尺寸图片幻灯片

图 4 本文方法的总体流程图

下载: 全尺寸图片幻灯片

图 5 本文方法的分类结果的混淆矩阵。

下载: 全尺寸图片幻灯片

表 1 VAIS数据集的训练样本和测试样本数量

序号	类名	训练	测试
1	medium-other	99	86
2	merchant	103	71
3	medium-passenger	78	62
4	sailing	214	198
5	small	342	313
6	tug	37	20
合计		873	750

下载: 导出CSV

表 2 自建数据集的训练样本和测试样本数量

序号	类名	训练	测试
1	散货船	1385	346
2	集装箱船	2381	595
3	客船	1632	408
4	帆船	1803	450
合计		7201	1799

下载: 导出CSV

表 3 不同方法在VAIS数据集上的分类准确率和对于每幅图像用于特征提取的平均时间消耗

方法	分类准确率(%)	特征提取时间(ms)
CNN1	92.13	0.104
CNN2	90.93	0.045
CNN3	90.67	0.092
本文方法	95.07	0.391

下载: 导出CSV

表 4 不同方法在自建数据集上的分类准确率和对于每幅图像用于特征提取的平均时间消耗

方法	分类准确率(%)	特征提取时间(ms)
CNN1	96.50	0.047
CNN2	94.61	0.048
CNN3	96.16	0.071
本文方法	97.50	0.396

下载: 导出CSV

表 5 不同方法在VAIS数据集上对每一类的分类准确率(%)

方法	类别
方法	medium-other	merchant	medium-passenger	sailing	small	tug
CNN1	80.23	87.32	88.71	94.95	94.89	100.00
CNN2	83.72	87.32	70.96	93.94	95.85	90.00
CNN3	81.40	94.37	75.80	100.00	89.78	85.00
本文方法	87.21	94.37	80.65	100.00	96.81	100.00

下载: 导出CSV

表 6 不同方法在自建数据集上对每一类的分类准确率(%)

方法	类别
方法	散货船	集装箱船	客船	帆船
CNN1	95.95	95.97	96.56	97.56
CNN2	89.60	96.30	96.81	94.22
CNN3	94.22	97.48	93.14	98.67
本文方法	96.53	97.82	96.57	98.67

下载: 导出CSV

表 7 本文方法与其他方法在VAIS数据集上的分类准确率和误分类样本数

方法	分类准确率(%)	误分类样本数
文献[3]方法	87.47	94
文献[20]方法	90.27	73
文献[13]方法	84.80	114
文献[12]方法	90.93	68
文献[21]方法	86.93	98
文献[22]方法	90.00	75
最大输出概率决策融合法	94.80	39
多数投票法	94.00	45
本文方法	95.07	37

下载: 导出CSV

表 8 本文方法与其他方法在自建数据集上的分类准确率和误分类样本数

方法	分类准确率(%)	误分类样本数
文献[3]方法	89.16	195
文献[20]方法	93.72	113
文献[12]方法	94.61	97
文献[21]方法	92.94	127
文献[22]方法	96.22	68
文献[13]方法	94.11	106
最大输出概率决策融合法	97.11	52
多数投票法	96.83	57
本文方法	97.50	45

下载: 导出CSV

参考文献(22)

[1]	ZHANG Erhu, WANG Kelu, and LIN Guangfeng. Classification of marine vessels with multi-feature structure fusion[J]. Applied Sciences, 2019, 9(10): 2153. doi: 10.3390/app9102153
[2]	DONG Chao, LIU Jinghong, and XU Fang. Ship detection in optical remote sensing images based on saliency and a rotation-invariant descriptor[J]. Remote Sensing, 2018, 10(3): 400. doi: 10.3390/rs10030400
[3]	吴映铮, 杨柳涛. 基于HOG和SVM的船舶图像分类算法[J]. 上海船舶运输科学研究所学报, 2019, 42(1): 58–64. WU Yingzheng and YANG Liutao. Ship image classification by combined use of HOG and SVM[J]. Journal of Shanghai Ship and Shipping Research Institute, 2019, 42(1): 58–64.
[4]	PARAMESWARAN S and RAINEY K. Vessel classification in overhead satellite imagery using weighted “bag of visual words”[C]. SPIE 9476, Automatic Target Recognition XXV, Baltimore, USA, 2015: 947609. doi: 10.1117/12.2177779.
[5]	ARGUEDAS V F. Texture-based vessel classifier for electro-optical satellite imagery[C]. 2015 IEEE International Conference on Image Processing, Quebec City, Canada, 2015: 3866–3870. doi: 10.1109/ICIP.2015.7351529.
[6]	王鑫, 李可, 宁晨, 等. 基于深度卷积神经网络和多核学习的遥感图像分类方法[J]. 电子与信息学报, 2019, 41(5): 1098–1105. doi: 10.11999/JEIT180628 WANG Xin, LI Ke, NING Chen, et al. Remote sensing image classification method based on deep convolution neural network and multi-kernel learning[J]. Journal of Electronics &Information Technology, 2019, 41(5): 1098–1105. doi: 10.11999/JEIT180628
[7]	李健伟, 曲长文, 彭书娟, 等. 基于生成对抗网络和线上难例挖掘的SAR图像舰船目标检测[J]. 电子与信息学报, 2019, 41(1): 143–149. doi: 10.11999/JEIT180050 LI Jianwei, QU Changwen, PENG Shujuan, et al. Ship detection in SAR images based on generative adversarial network and online hard examples mining[J]. Journal of Electronics &Information Technology, 2019, 41(1): 143–149. doi: 10.11999/JEIT180050
[8]	CHEN Yunfan, XIE Han, and SHIN H. Multi-layer fusion techniques using a CNN for multispectral pedestrian detection[J]. IET Computer Vision, 2018, 12(8): 1179–1187. doi: 10.1049/iet-cvi.2018.5315
[9]	闫河, 王鹏, 董莺艳, 等. 一种CNN与ELM相结合的船舶分类识别方法[J]. 重庆理工大学学报: 自然科学, 2019, 33(1): 53–57. doi: 10.3969/j.issn.1674-8425(z).2019.01.008 YAN He, WANG Peng, DONG Yingyan, et al. A classification identification method of ships combining CNN and ELM[J]. Journal of Chongqing Institute of Technology:Natural Science, 2019, 33(1): 53–57. doi: 10.3969/j.issn.1674-8425(z).2019.01.008
[10]	陈兴伟. 深度学习船舶分类技术研究[J]. 舰船科学技术, 2019, 41(7A): 142–144. doi: 10.3404/j.issn.1672-7649.2019.7A.048 CHEN Xingwei. Research on ship classification technology based on deep learning[J]. Ship Science and Technology, 2019, 41(7A): 142–144. doi: 10.3404/j.issn.1672-7649.2019.7A.048
[11]	SHI Qiaoqiao, LI Wei, TAO Ran, et al. Ship classification based on multifeature ensemble with convolutional neural network[J]. Remote Sensing, 2019, 11(4): 419. doi: 10.3390/rs11040419
[12]	BENTES C, VELOTTO D, and TINGS B. Ship classification in TerraSAR-X images with convolutional neural networks[J]. IEEE Journal of Oceanic Engineering, 2018, 43(1): 258–266. doi: 10.1109/JOE.2017.2767106
[13]	杨亚东, 王晓峰, 潘静静. 改进CNN及其在船舶识别中的应用[J]. 计算机工程与设计, 2018, 39(10): 3228–3233. doi: 10.16208/j.issn1000-7024.2018.10.039 YANG Yadong, WANG Xiaofeng, and PAN Jingjing. Improved CNN and its application in ship identification[J]. Computer Engineering and Design, 2018, 39(10): 3228–3233. doi: 10.16208/j.issn1000-7024.2018.10.039
[14]	LI Xiaobin, JIANG Bitao, SUN Tong, et al. Remote sensing scene classification based on decision-level fusion[C]. 2018 IEEE 4th Information Technology and Mechatronics Engineering Conference, Chongqing, China, 2018: 393–397. doi: 10.1109/ITOEC.2018.8740526.
[15]	GENG Jie, JIANG Wen, and DENG Xinyang. Multi-scale deep feature learning network with bilateral filtering for SAR image classification[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2020, 167: 201–213. doi: 10.1016/j.isprsjprs.2020.07.007
[16]	李凯, 韩冰, 张景滔. 基于条件随机场与多尺度卷积神经网络的交通标志检测与识别[J]. 计算机应用, 2018, 38(S2): 270–275. LI Kai, HAN Bing, and ZHANG Jingtao. Traffic sign detection and recognition based on conditional random field and multi-scale convolutional neural network[J]. Journal of Computer Applications, 2018, 38(S2): 270–275.
[17]	REN Yongmei, YANG Jie, ZHANG Qingnian, et al. Multi-feature fusion with convolutional neural network for ship classification in optical images[J]. Applied Sciences, 2019, 9(20): 4209. doi: 10.3390/app9204209
[18]	CHEN Wangcai, LIU Wenbo and LI Kaiyu. Rail crack recognition based on adaptive weighting multi-classifier fusion decision[J]. Measurement, 2018, 123: 102–114. doi: 10.1016/j.measurement.2018.03.059
[19]	ZHANG M M, CHOI J, DANIILIDIS K, et al. VAIS: A dataset for recognizing maritime imagery in the visible and infrared spectrums[C]. The 2015 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Boston, USA, 2015: 10–16. doi: 10.1109/CVPRW.2015.7301291.
[20]	DING Jun, CHEN Bo, LIU Hongwei, et al. Convolutional neural network with data augmentation for SAR target recognition[J]. IEEE Geoscience and Remote Sensing Letters, 2016, 13(3): 364–368. doi: 10.1109/LGRS.2015.2513754
[21]	RAINEY K, REEDER J D, and CORELLI A G. Convolution neural networks for ship type recognition[C]. SPIE 9844, Automatic Target Recognition XXVI, Baltimore, USA, 2016: 984409. doi: 10.1117/12.2229366.
[22]	LI Zhenzhen, ZHAO Baojun, TANG Linbo, et al. Ship classification based on convolutional neural networks[J]. The Journal of Engineering, 2019, 2019(21): 7343–7346. doi: 10.1049/joe.2019.0422