Multi-Target Detection of Low-Illuminance Scene Based on Polarization Image
-
摘要: 偏振光反射信息可直接反演目标本征特性,且在传输过程中具备较强的抗干扰特性,因此偏振成像技术可适用于多种复杂环境中的智能监控、交通监察领域。近年来使用深度学习判读图像检测目标的方法迅速发展,已经广泛应用于图像处理的各个领域。本文提出了一种基于偏振图像与深度神经网络算法的行人、车辆多目标检测算法YOLOv5s-DOLP。首先,通过实时获取到偏振图像进行偏振信息解析,获取目标偏振度图像。其次,为增强偏振度图像中检测目标与背景存在高对比度的特性,在主干网络中引入通道注意力与空间注意力,提升网络特征进行自适应学习的能力。此外,使用K-means算法对目标位置信息进行聚类分析,加快网络在偏振度图像的学习速度,提升目标检测精度。实验结果显示,该算法结合了偏振成像和深度学习目标检测的优势,对于低照度复杂场景中的车辆、行人目标检测效果好、检测速度快,对于道路车辆的目标检测、识别与跟踪具有一定的应用价值。Abstract: Polarized light reflection information can directly invert the intrinsic characteristics of a target and has strong anti-interference characteristics in the transmission process. Thus, polarization imaging technology can be applied to the fields of intelligent monitoring and traffic monitoring in various complex environments. In recent years, deep-neural-network methods for interpreting image detection targets have been developed rapidly and widely used in various fields of image processing. In this study, a vehicle multi-target detection algorithm based on polarized images and deep learning is proposed. First, the target polarization degree image can be obtained by acquiring the polarization image in real time and analyzing the polarization information. Second, to enhance the high contrast between the detection targets and the background in the polarization image, channel attention and spatial attention are introduced into the backbone network to improve the ability of the network features to perform adaptive learning. In addition, the K-means algorithm is used to perform clustering analysis on the target location information, thereby increasing the network's learning speed in the polarization image and improving the progress of target detection. The experimental results show that this method is effective and fast for vehicle detection in complex scenes with low illumination. This method combines the advantages of polarization imaging and deep-learning target detection and has substantial application scope in road vehicle target detection, recognition, and tracking.
-
Key words:
- polarization image /
- neural network /
- YOLO v5s /
- multi-target detection /
- attention mechanism
-
图 5 拉普拉斯算子[15]检测结果
Figure 5. Detecting result of Laplace
表 1 分光型偏振成像设备
Table 1. Spectroscopic polarization imaging equipment
Max resolution Frame Image sensor Pixel
size/μmADC 2488×2048 65 Sony IMX250,CMOS,2/3 Global shutter 3.45 10 bit/12 bit 表 2 网络训练环境
Table 2. Network training environment
Name Configure CPU Intel Xeon E5-2630 GPU NVIDIA 1080Ti * 2 Operating system Ubuntu 18.04 Parallel computing library Cuda10 + Cudnn7.4 Image processing Python3.6、Opencv3.4.0 Deep learning framework Pytorch 表 3 目标检测算法检测结果
Table 3. Detect results of target detection algorithm
Methods AP mAP Test time/s Person Car Faster R-CNN(Res 50) 91.6 97.5 94.6 0.103 YOLOv4(Res 50) 91.6 97.3 94.5 0.029 YOLOv5s-DOLP 94.8 98.9 96.9 0.039 表 4 消融实验结果
Table 4. Results of ablation experiment
Methods AP mAP Test time/s Person Car YOLOv5s 87.2 98.3 92.8 0.027 A 90.2 98.5 94.3 0.029 B 92.0 98.8 95.4 0.034 C 88.5 98.4 93.4 0.027 -
[1] 于洵, 杨烨, 姜旭, 等. 基于偏振光谱成像的目标识别方法研究[J]. 应用光学, 2016, 37(4): 537-541. https://www.cnki.com.cn/Article/CJFDTOTAL-YYGX201604008.htmYU Xun, YANG Ye, JIANG Xu, et al. Recognition of camouflage targets by polarization spectral imaging system[J]. Journal of Applied Optics, 2016, 37(4): 537-541. https://www.cnki.com.cn/Article/CJFDTOTAL-YYGX201604008.htm [2] 李从利, 薛松, 陆文骏, 等. 雾天条件下偏振解析成像质量评价[J]. 中国图象图形学报, 2017, 22(3): 366-375. https://www.cnki.com.cn/Article/CJFDTOTAL-ZGTB201703011.htmLI Congli, XUE Song, LU Wenjun, et al. Quality assessment of polarization imaging under foggy[J]. Journal of Image and Graphics, 2017, 22(3): 366-375. https://www.cnki.com.cn/Article/CJFDTOTAL-ZGTB201703011.htm [3] 李小明, 黄勤超. 沙漠背景下红外偏振图像目标检测方法[J]. 红外技术, 2016, 38(9): 779-782, 792. http://hwjs.nvir.cn/article/id/hwjs201609012LI Xiaoming, HUANG Qinchao. Target detection for infrared polarization image in the background of desert[J]. Infrared Technology, 2016, 38(9): 779-782, 792. http://hwjs.nvir.cn/article/id/hwjs201609012 [4] 张荣, 李伟平, 莫同. 深度学习研究综述[J]. 信息与控制, 2018, 47(4): 385-397. https://www.cnki.com.cn/Article/CJFDTOTAL-XXYK201804002.htmZHANG Rong, LI Weiping, MO Tong. Review of deep learning[J]. Information and Control, 2018, 47(4): 385-397. https://www.cnki.com.cn/Article/CJFDTOTAL-XXYK201804002.htm [5] 王文秀, 傅雨田, 董峰, 等. 基于深度卷积神经网络的红外船只目标检测方法[J]. 光学学报, 2018, 38(7): 0712006. https://www.cnki.com.cn/Article/CJFDTOTAL-GXXB201807020.htmWANG Wenxiu, FU Yutian, DONG Feng, et al. Infrared ship target detection method based on deep convolution neural network[J]. Acta Optica Sinica, 2018, 38(7): 0712006. https://www.cnki.com.cn/Article/CJFDTOTAL-GXXB201807020.htm [6] 罗海波, 许凌云, 惠斌, 等. 基于深度学习的目标跟踪方法研究现状与展望[J]. 红外与激光工程, 2017, 46(5): 0502002-0502002(7). https://www.cnki.com.cn/Article/CJFDTOTAL-HWYJ201705002.htmLUO Haibo, XU Lingyun, HUI Bin, et al. Status and prospect of target tracking based on deep learning[J]. Infrared and Laser Engineering, 2017, 46(5): 0502002-0502002(7). https://www.cnki.com.cn/Article/CJFDTOTAL-HWYJ201705002.htm [7] Ross Girshick. Fast R-CNN[J]. Computer Science, 2015, 6: 1440-1448. [8] REN Shaoqing, HE Kaiming, Ross Girshick, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(6): 1137-1149. [9] HE Kaiming, Georgia Gkioxari, Piotr Dollar, et al. Mask R-CNN[C]// Proceedings of the IEEE Computer Society Conference on Computer Vision, 2017: 2961-2969. [10] LIU Wei, Dragomir Anguelov, Dumitru Erhan, et al. SSD: Single shot multibox detector[C]//European Conference on Computer Vision, 2016, 6: 21-27. [11] Joseph Redmon, Santosh Divvala, Ross Girshick, et al. You Only Look Once: unified, real-time object detection[C]//Conference on Computer Vision and Pattern Recognition, 2016, 6: 779-788. [12] Joseph Redmon, Ali Farhadi. YOLO9000: better, faster, stronger[C]// Conference on Computer Vision and Pattern Recognition, 2017, 7: 6517-6525 [13] Joseph Redmon, Ali Farhadi. YOLOv3: an incremental improvement [J/OL]. arXiv: 1804.02767, 2018. [14] Alexey Bochkovskiy, Chien-Yao Wang, Hong-Yuan Mark Liao. YOLOv4: optimal speed and accuracy of object detection[J/OL]. arXiv: 2004. 10934, 2020. [15] 冈萨雷斯. 数字图像处理[M]. 第三版, 北京: 电子工业出版社, 2011: 26-29.Rafael C. Gonzalez. Digital Image Processing[M]. 3th, Beijing: Publishing House of Electronics Industry, 2011: 26-29. [16] Woo S, Park J, Lee J Y, et al. Cbam: Convolutional block attention module[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 3-19. [17] 张汝榛, 张建林, 祁小平, 等. 复杂场景下的红外目标检测[J]. 光电工程, 2020, 47(10): 200314. https://www.cnki.com.cn/Article/CJFDTOTAL-GDGC202010010.htmZHANG R Z, ZHANG J L, QI X P, et al. Infrared target detection and recognition in complex scene[J]. Opto-Electron Eng., 2020, 47(10): 200314. https://www.cnki.com.cn/Article/CJFDTOTAL-GDGC202010010.htm [18] 宫剑, 吕俊伟, 刘亮, 等. 红外偏振图像的舰船目标检测[J]. 光谱学与光谱分析, 2020, 40(2): 586-594. https://www.cnki.com.cn/Article/CJFDTOTAL-GUAN202002052.htmGONG Jian, LYU Junwei, LIU Liang, et al. Ship target detection based on infrared polarization image[J]. Spectroscopy and Spectral Analysis, 2020, 40(2): 586-594. https://www.cnki.com.cn/Article/CJFDTOTAL-GUAN202002052.htm [19] 游江, 刘鹏祖, 容晓龙, 等. 基于暗通道先验原理的偏振图像去雾增强算法研究[J]. 激光与红外, 2020, 50(4): 493-500. doi: 10.3969/j.issn.1001-5078.2020.04.019YOU Jiang, LIU Pengzu, RONG Xiaolong, et al. Dehazing and enhancement research of polarized image based on dark channel priori principle[J]. Laser & Infrared, 2020, 50(4): 493-500. doi: 10.3969/j.issn.1001-5078.2020.04.019 [20] 王美荣, 徐国明, 袁宏武. 显著性偏振参量深度稀疏特征学习的目标检测方法[J]. 激光与光电子学进展, 2019, 56(19): 191101. https://www.cnki.com.cn/Article/CJFDTOTAL-JGDJ201919013.htmWANG Meirong, XU Guoming, YUAN Hongwu. Object detection by deep sparse feature learning of salient polarization parameters[J]. Laser & Optoelectronics Progress, 2019, 56(19): 191101. https://www.cnki.com.cn/Article/CJFDTOTAL-JGDJ201919013.htm [21] 李慕锴, 张涛, 崔文楠. 基于YOLOv3的红外行人小目标检测技术研究[J]. 红外技术, 2020, 42(2): 176-181. http://hwjs.nvir.cn/article/id/hwjs202002012LI Mukai, ZHANG Tao, CUI Wennan. Research of infrared small pedestrian target detection based on YOLOv3[J]. Infrared Technology, 2020, 42(2): 176-181. http://hwjs.nvir.cn/article/id/hwjs202002012