支持抗光照目标检测的改进YOLO算法

doi:10.19562/j.chinasae.qcgc.2023.05.007

摘要/Abstract

摘要：

针对现有的深度学习目标检测算法中存在的复杂光照场景下检测精度不高、实时性差等问题，提出了一种基于YOLO算法的抗光照目标检测网络模型YOLO-RLG。首先，将输入模型的RGB数据转换为HSV数据，从HSV数据分离出抗光照能力强的S通道，并与RGB数据合并生成RGBS数据，使输入数据具备抗光照能力；其次，将YOLOV4的主干网络替换成Ghostnet网络，并对其在普通卷积与廉价卷积的模型分配比例上进行调整，在保证检测精度的同时提高检测速度；最后，用EIoU替换CIoU改进模型的损失函数，提高了目标检测精度和算法鲁棒性。基于KITTI与VOC数据集的实验结果表明，与原网络模型比较，FPS提高了22.54与17.84 f/s，模型降低了210.3 M，精确度（AP）提升了0.83%与1.31%，且算法的抗光照能力得到显著增强。

关键词: 机器视觉, 抗光照图像处理, Ghostnet网络, 损失函数

Abstract:

For the problems of unsatisfactory detection accuracy and weak real-time performance in the complicated illumination scenes in the existing deep learning target detection algorithms， an anti-illumination target detection network model YOLO-RLG based on the YOLO algorithm is proposed. Firstly， the RGB data of the input model is converted into HSV data， and the S channel with powerful anti-illumination capability is separated from the HSV data and fused with the RGB data to generate RGBS data so that the input data has anti-illumination capability. Secondly， the backbone network of YOLOV4 is replaced with Ghostnet network， with the model assignment ratio between ordinary convolution and cheap convolution modified to improve the detection efficiency while ensuring the detection accuracy. Finally， the loss function of the model is improved by replacing CIoU with EIoU， which enhances the target detection accuracy and algorithm robustness. The experimental results based on KITTI and VOC datasets indicate that， compared with the original network model， the FPS improves by 22.54 and 17.84 f/s， with the model reduced by 210.3 M， the accuracy （AP） improved by 0.83% and 1.31%， and the algorithm's anti-illumination performance significantly enhanced.

Key words: machine vision, anti-illumination image processing, Ghostnet network, loss function

姚宇捷,彭育辉,陈泽辉,何维堃,吴庆,黄炜,陈文强. 支持抗光照目标检测的改进YOLO算法[J]. 汽车工程, 2023, 45(5): 777-785.

Yujie Yao,Yuhui Peng,Zehui Chen,Weikun He,Qing Wu,Wei Huang,Wenqiang Chen. An Improved YOLO Algorithm Supporting Anti-illumination Target Detection[J]. Automotive Engineering, 2023, 45(5): 777-785.

图/表 13

图1

图2

图3

图4

图5

图6

图7

表1

图8

表2

图9

图10

表3

参考文献 23

1	CIELNIAK G， DUCKETT T， LILIENTHAL A J. Data association and occlusion handling for vision-based people tracking by mobile robots［J］. Robotics and Autonomous Systems， 2010， 58（5）： 435-443.
2	SATAKE J， CHIBA M， MIURA J. Visual person identification using a distance-dependent appearance model for a person following robot［J］. International Journal of Automation and Computing， 2013， 10（5）： 438-446.
3	GIRSHICK R，DONAHUE J，DARRELL T，et al. Rich feature hierarchies for accurate object detection and semantic segmentation［C］. Proceedings of 2014 IEEE Conference on Computer Vision and Pattern Recognition， 2014：580-587.
4	REDMON J， DIVVALA S， GIRSHICK R， et al. You only look once： unified， real-time object detection［C］. Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition， 2016： 779-788.
5	REDMON J， FARHADI A. YOLO9000： better， faster， stronger［C］. Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition， 2017： 7263-7271.
6	REDMON J， FARHADI A. Yolov3： an incremental improvement［J］. arXiv preprint arXiv：， 2018.
7	LIU W， ANGUELOV D， ERHAN D， et al. SSD： single shot multibox detector［C］. European Conference on Computer Vision. Springer， Cham， 2016： 21-37.
8	HOWARD A G， ZHU M， CHEN B， et al. Mobilenets： efficient convolutional neural networks for mobile vision applications［J］. arXiv preprint arXiv：， 2017.
9	SANDLER M， HOWARD A， ZHU M， et al. Mobilenetv2： inverted residuals and linear bottlenecks［C］. Proceedings of 2018 IEEE Conference on Computer Vision and Pattern Recognition， 2018： 4510-4520.
10	HOWARD A， SANDLER M， CHU G， et al. Searching for mobilenetv3［C］. Proceedings of 2019 IEEE/CVF International Conference on Computer Vision， 2019： 1314-1324.
11	REN J， CHEN X， LIU J， et al. Accurate single stage detector using recurrent rolling convolution［C］. Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition， 2017： 5420-5428.
12	ZHAO M， ZHONG Y， SUN D， et al. Accurate and efficient vehicle detection framework based on SSD algorithm［J］. IET Image Processing， 2021， 15（13）： 3094-3104.
13	ZHU Y W， ZHOU H， WANG Y， et al. Research on fast detection algorithm of traffic road target［C］. International Conference on High Performance Computing and Communication （HPCCE 2021）. SPIE， 2022， 12162： 330-337.
14	LIU S， QI L， QIN H， et al. Path aggregation network for instance segmentation［C］. Proceedings of 2018 IEEE Conference on Computer Vision and Pattern Recognition， 2018： 8759-8768.
15	BOCHKOVSKIY A， WANG C Y， LIAO H Y M. Yolov4： optimal speed and accuracy of object detection［J］. arXiv preprint arXiv：， 2020.
16	HAN K， WANG Y， TIAN Q， et al. Ghostnet： more features from cheap operations［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2020： 1580-1589.
17	ZHANG Y F， REN W， ZHANG Z， et al. Focal and efficient IOU loss for accurate bounding box regression［J］. arXiv preprint arXiv：， 2021.
18	REN S， HE K， GIRSHICK R， et al. Faster r-cnn： towards real-time object detection with region proposal networks［J］. Advances in Neural Information Processing Systems， 2015， 28： 91-99.
19	NANDIMANDALAM V D. Military and non-military vehicle detection by faster R-CNN and SSD300 models using transfer leaning［D］. Dublin： National College of Ireland， 2020.
20	JEONG J， PARK H， KWAK N. Enhancement of SSD by concatenating feature maps for object detection［J］. arXiv preprint arXiv：， 2017.
21	廖慕钦，周永军，汤小红，等.基于SSD-MobilenetV3模型的车辆检测［J］.传感器与微系统，2022，41（6）：142-145.
	LIAO M Q， ZHOU Y J， TANG X H， et al. Vehicle detection based on SSD-MobilenetV3 model［J］.Sensors and Microsystems， 2022，41（6）：142-145.
22	TAN M， PANG R， LE Q V. Efficientdet： scalable and efficient object detection［C］. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition， 2020： 10781-10790.
23	LIU Z， LIN Y， CAO Y， et al. Swin transformer： hierarchical vision transformer using shifted windows［C］. Proceedings of the IEEE/CVF International Conference on Computer Vision， 2021： 10012-10022.

模型	抗光照增强	Ghostnet替换	Ghostnet改进	EIoU	AP/%	MS/M	FPS/（f·s^-1）
YOLOV4	×	×	×	×	92.53	256.8	34.04
Model 1	√	×	×	×	94.08	263.7	34.22
Model 2	√	√	×	×	88.75	45.4	52.63
Model 3	√	√	√	×	92.62	46.5	55.91
YOLO-RLG	√	√	√	√	93.36	46.5	56.58

模型	AP/%	MS/M	FPS/（f·s^-1）
Faster-RCNN^［18］	77.09	522.4	12.47
SSD300^［19］	81.59	50.6	56.71
SSD512^［20］	84.81	100.2	26.24
SSD-MoblienetV3^［21］	85.6	16.9	55.32
YOLOV4	92.53	256.8	34.04
YOLOV5	90.93	180.5	38.58
EfficientDet^［22］	91.57	355.0	15.50
Swin Transformer^［23］	96.34	1 149.3	29.26
YOLO-RLG	93.36	46.5	56.58

模型	mAP/%	Size/M	FPS/（f·s^-1）
YOLOV4	91.33	256.8	27.04
YOLO-RLG	92.64	46.5	44.88

[1]	程腾,倪昊,张强,王文冲,石琴. 基于虚拟点云的二阶段多模态融合网络[J]. 汽车工程, 2024, 46(2): 222-229.
[2]	石琴,陈雅芳,程腾,张强,王文冲,石本义. 基于局部平面引导层的无监督单目红外图像深度估计[J]. 汽车工程, 2023, 45(12): 2291-2298.
[3]	张炳力,秦浩然,江尚,郑杰禹,吴正海. 基于RetinaNet及优化损失函数的夜间车辆检测方法[J]. 汽车工程, 2021, 43(8): 1195-1202.
[4]	朱烨添, 沈鸿, 张文俊, 林立, 吴少腾, 敖云轲, 万禹, 赵之谦, 裴景玉. 汽车自动变速器阀体阀芯孔内壁面质量检测系统的研发^*[J]. 汽车工程, 2020, 42(4): 484-490.
[5]	白中浩, 李智强, 蒋彬辉, 王鹏辉. 基于改进YOLOv2模型的驾驶辅助系统实时行人检测^*[J]. 汽车工程, 2019, 41(12): 1416-1423.
[6]	王战古,高松,邵金菊，谭德荣，孙亮，于杰. 基于深度置信网络的多源信息前方车辆检测[J]. 汽车工程, 2018, 40(5): 554-560.