基于自适应特征融合和注意力机制的变电设备红外图像识别

王媛彬; 吴冰超

doi:10.11999/JEIT231047

基于自适应特征融合和注意力机制的变电设备红外图像识别

doi: 10.11999/JEIT231047

王媛彬^,,
吴冰超

1.
西安科技大学电气与控制工程学院西安 710054
2.
西安市电气设备状态监测与供电安全重点实验室西安 710054

基金项目: 国家自然科学基金(52174198)，陕西省重点研发计划(2023 YBSF-133)，陕西省秦创原科学家+工程师队伍建设(2022KXJ-166)

详细信息

作者简介:
王媛彬：女，博士，副教授，研究方向为电力设备状态检测

吴冰超：男，硕士生，研究方向为电力设备图像处理

通讯作者:
王媛彬　wangyb998@163.com

中图分类号: TN911.73
计量
- 文章访问数: 275
- HTML全文浏览量: 172
- PDF下载量: 50
- 被引次数: 0
出版历程
- 收稿日期: 2023-09-27
- 修回日期: 2023-12-03
- 网络出版日期: 2023-12-14
- 刊出日期: 2024-09-26

Infrared Image Recognition of Substation Equipment Based on Adaptive Feature Fusion and Attention Mechanism

WANG Yuanbin^,,
WU Bingchao

1.
College of Electrical and Control Engineering, Xi’an University of Science and Technology, Xi’an 710054, China
2.
Xi’an Key Laboratory of Electrical Equipment Condition Monitoring and Power Supply Security, Xi’an 710054, China

Funds: The National Natural Science Foundation of China (52174198), The National Key Research and Development Program of Shaanxi Province (2023 YBSF-133), The Construction of Qinchuangyuan Scientists and Engineers Team in Shaanxi Province (2022KXJ-166)

摘要

摘要: 针对变电设备红外图像复杂背景下多目标、小目标及遮挡目标识别效果差的问题，该文提出一种基于中心点网络(CenterNet)的变电设备红外图像识别方法。通过将自适应特征融合模块(ASFF)和特征金字塔(FPN)相结合, 构建ASFF+FPN结构的特征融合网络，增强了模型对多目标和小目标的跨尺度特征融合能力，排除背景信息；针对网络对遮挡目标特征捕捉能力差的问题，在特征融合网络中添加全局注意力机制，增强目标显著度；为实现模型轻量化，引入深度可分离卷积，减少参数量和推理时间；最后，通过引入分布焦点损失函数，克服了原损失函数对遮挡目标敏感性差的问题，提升了模型收敛速度和识别精度。在包含7种红外变电设备图像的自建数据集上进行测试。实验表明该算法与原始算法相比，识别精度提升了3.55%，达到了95.19%，模型参数量仅为32.52M，与4种主流目标识别算法对比，该算法在识别精度和算法复杂度上具有明显优势。
- 变电设备 /
- 红外图像识别 /
- 中心点网络 /
- 自适应特征融合 /
- 注意力机制
Abstract: To address the challenges of poor recognition effect of the infrared substation equipment image caused by multi-target, small target and occlusion target in complex background, an infrared image recognition method for substation equipment based on CenterNet is proposed. By combining the Adaptive Spatial Feature Fusion(ASFF) module and Feature Pyramid Networks (FPN), a feature fusion network with the structure of ASFF+FPN is constructed, and the cross-scale feature fusion capability of the model for multi-target and small target is enhanced, which excludes background information. To improve the feature capturing ability of occluding targets, the global attention mechanism is introduced to the feature fusion network to enhance target saliency. Additionally, depthwise separable convolution is introduced to reduce parameters number and model inference time, and a lightweight model is achieved. Finally, the problem of poor sensitivity to obscured targets is overcame by using the distribution focal loss function, and the convergence speed and recognition accuracy is improved. Tests are performed on a self-built dataset containing seven infrared substation equipment images. Experimental results demonstrate that the proposed algorithm achieves a recognition accuracy of 95.19%, an improvement of 3.55% compared with the original algorithm, while it only has 32.52M model parameters. Furthermore, the method shows significant advantages in recognition accuracy and algorithm complexity, over four main target recognition algorithms.
- Substation equipment /
- Infrared image recognition /
- CenterNet /
- Adaptive Spatial Feature Fusion (ASFF) /
- Attention mechanism.

HTML全文

图 1 CenterNet网络结构

下载: 全尺寸图片幻灯片

图 2 本文算法结构框架

下载: 全尺寸图片幻灯片

图 3 ASFF原理

下载: 全尺寸图片幻灯片

图 4 GAM工作原理

下载: 全尺寸图片幻灯片

图 5 不同λ_size,λ_off取值的识别精度

下载: 全尺寸图片幻灯片

图 6 评价指标变化曲线

下载: 全尺寸图片幻灯片

图 7 识别效果对比图

下载: 全尺寸图片幻灯片

表 1 不同设备识别结果(%)

设备类型	精确率	召回率	mAP0.5
电流互感器	97.84	90.07	98.53
电压互感器	80.00	81.36	88.60
隔离开关	96.26	99.45	99.58
绝缘子	96.58	89.40	95.56
避雷器	95.88	90.73	95.15
套管	92.37	81.34	89.80
断路器	99.49	91.63	99.09
所有种类	94.06	89.14	95.19

下载: 导出CSV

表 2 消融实验

方法	ASFF+FPN	GAM	DSC	DFL	mAP0.5(%)
CenterNet	×	×	×	×	91.64
方案1	√	×	×	×	92.43
方案2	√	√	×	×	93.10
方案3	√	√	√	×	93.32
本文方法	√	√	√	√	95.19

下载: 导出CSV

表 3 不同算法对比实验

算法	mAP0.5 (%)	FPS	参数量(M)	FLOPs(G)
Faster R-CNN	94.65	11.84	137.10	370.21
SSD	91.92	76.73	36.28	72.75
YOLOv5	91.50	87.50	47.06	115.91
YOLOv7	88.04	36.54	37.62	106.47
CenterNet	91.64	61.93	32.67	70.21
本文算法	95.19	65.32	32.52	69.28

下载: 导出CSV

表 4 泛化性实验mAP0.5结果(%)

识别类型	缺陷绝缘子	正常绝缘子	所有种类
CenterNet	90.26	92.53	91.40
本文算法	93.69	96.45	95.07

下载: 导出CSV

参考文献(19)

[1]	LI Jianqi, XU Yaqian, NIE Keheng, et al. PEDNet: A lightweight detection network of power equipment in infrared image based on YOLOv4-Tiny[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 5004312. doi: 10.1109/tim.2023.3235416.
[2]	XIA Changjie, REN Ming, WANG Bing, et al. Infrared thermography-based diagnostics on power equipment: State-of-the-art[J]. High Voltage, 2021, 6(3): 387–407. doi: 10.1049/hve2.12023.
[3]	REN Shaoqing, HE Kaiming, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal network[C]. The 28th International Conference on Neural Information Processing Systems, Montreal, Canada, 2015: 91–99.
[4]	OU Jianhua, WANG Jianguo, XUE Jian, et al. Infrared image target detection of substation electrical equipment using an improved faster R-CNN[J]. IEEE Transactions on Power Delivery, 2023, 38(1): 387–396. doi: 10.1109/tpwrd.2022.3191694.
[5]	王妤, 陈秀新, 袁和金. 基于改进Faster RCNN的变电站红外图像多目标识别[J]. 传感技术学报, 2021, 34(4): 522–530. doi: 10.3969/j.issn.1004-1699.2021.04.014. WANG Yu, CHEN Xiuxin, and YUAN Hejin. Multi-target recognition of substation infrared image based on improved Faster RCNN[J]. Chinese Journal of Sensors and Actuators, 2021, 34(4): 522–530. doi: 10.3969/j.issn.1004-1699.2021.04.014.
[6]	LIU Wei, ANGUELOV D, ERHAN D, et al. SSD: Single shot MultiBox detector[C]. The 14th European Conference on Computer Vision, Amsterdam, The Netherlands, 2016: 21–37. doi: 10.1007/978-3-319-46448-0_2.
[7]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 779–788. doi: 10.1109/CVPR.2016.91.
[8]	王旭红, 李浩, 樊绍胜, 等. 基于改进SSD的电力设备红外图像异常自动检测方法[J]. 电工技术学报, 2020, 35(S1): 302–310. doi: 10.19595/j.cnki.1000-6753.tces.l80426. WANG Xuhong, LI Hao, FAN Shaosheng, et al. Infrared image anomaly automatic detection method for power equipment based on improved single shot multi box detection[J]. Transactions of China Electrotechnical Society, 2020, 35(S1): 302–310. doi: 10.19595/j.cnki.1000-6753.tces.l80426.
[9]	谭宇璇, 樊绍胜. 基于图像增强与深度学习的变电设备红外热像识别方法[J]. 中国电机工程学报, 2021, 41(23): 7990–7997. doi: 10.13334/j.0258-8013.pcsee.202487. TAN Yuxuan and FAN Shaosheng. Infrared thermal image recognition of substation equipment based on image enhancement and deep learning[J]. Proceedings of the CSEE, 2021, 41(23): 7990–7997. doi: 10.13334/j.0258-8013.pcsee.202487.
[10]	王媛彬, 李媛媛, 段誉, 等. 基于轻量骨干网络和注意力结构的变电设备红外图像识别[J]. 电网技术, 2023, 47(10): 4358–4369. doi: 10.13335/j.1000-3673.pst.2022.2113. WANG Yuanbin, LI Yuanyuan, DUAN Yu, et al. Infrared image recognition of substation equipment based on lightweight backbone network and attention mechanism[J]. Power System Technology, 2023, 47(10): 4358–4369. doi: 10.13335/j.1000-3673.pst.2022.2113.
[11]	DUAN Kaiwen, BAI Song, XIE Lingxi, et al. CenterNet: Keypoint triplets for object detection[C]. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea (South), 2019: 6568–6577. doi: 10.1109/iccv.2019.00667.
[12]	黄悦华, 杨楚睿, 陈晨, 等. 基于改进Centernet的变电设备红外检测方法[J]. 电子测量技术, 2023, 46(4): 142–148. doi: 10.19651/j.cnki.emt.2210406. HUANG Yuehua, YANG Churui, CHEN Chen, et al. Infrared detection method of substation equipment based on improved Centernet[J]. Electronic Measurement Technology, 2023, 46(4): 142–148. doi: 10.19651/j.cnki.emt.2210406.
[13]	ZHENG Hanbo, CUI Yaohui, FAN Xianhao, et al. An infrared image detection method of substation equipment combining Iresgroup structure and CenterNet[J]. IEEE Transactions on Power Delivery, 2022, 37(6): 4757–4765. doi: 10.1109/tpwrd.2022.3158818.
[14]	CHENG Xun and YU Jianbo. RetinaNet with difference channel attention and adaptively spatial feature fusion for steel surface defect detection[J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 2503911. doi: 10.1109/tim.2020.3040485.
[15]	LI Yuancheng, ZHOU Shenglong, CHEN Hui, et al. Attention-based fusion factor in FPN for object detection[J]. Applied Intelligence, 2022, 52(13): 15547–15556. doi: 10.1007/s10489-022-03220-0.
[16]	NIU Zhaoyang, ZHONG Guoqiang, and YU Hui. A review on the attention mechanism of deep learning[J]. Neurocomputing, 2021, 452: 48–62. doi: 10.1016/j.neucom.2021.03.091.
[17]	谢雯, 王若男, 羊鑫, 等. 融合深度可分离卷积的多尺度残差UNet在PolSAR地物分类中的研究[J]. 电子与信息学报, 2023, 45(8): 2975–2985. doi: 10.11999/JEIT220867. XIE Wen, WANG Ruonan, YANG Xin, et al. Research on multi-scale residual UNet fused with depthwise separable convolution in PolSAR terrain classification[J]. Journal of Electronics & Information Technology, 2023, 45(8): 2975–2985. doi: 10.11999/JEIT220867.
[18]	LI Feng, ZURADA J M, and WU Wei. Smooth group L_1/2 regularization for input layer of feedforward neural networks[J]. Neurocomputing, 2018, 314: 109–119. doi: 10.1016/j.neucom.2018.06.046.
[19]	康守强, 杨佳轩, 王玉静, 等. 基于改进宽度模型迁移学习的不同负载下滚动轴承状态快速分类方法[J]. 电子与信息学报, 2023, 45(5): 1824–1832. doi: 10.11999/JEIT220401. KANG Shouqiang, YANG Jiaxuan, WANG Yujing, et al. A fast classification method of rolling bearing state under different loads based on improved broad model transfer learning[J]. Journal of Electronics & Information Technology, 2023, 45(5): 1824–1832. doi: 10.11999/JEIT 220401.