应用归一化通道注意力机制的YOLOv7交通标志检测算法

doi:10.3778/j.issn.1002-8331.2403-0084

摘要/Abstract

摘要： 现有目标检测算法对背景复杂下小交通标志的检测效果并不理想。为此，提出了一种基于归一化通道注意力机制YOLOv7的交通标志检测算法（YOLOv7 based on normalized channel attention mechanism，YOLOv7- NCAM）。为了使YOLOv7-NCAM模型具有像素级建模能力，提高它对小目标交通标志特征的提取能力，YOLOv7-NCAM算法使用FReLU激活函数构建了DBF和CBF两种卷积层，并用它们来组建模型的Backbone模块和Neck模块；提出一种归一化通道注意力机制（normalized channel attention mechanism，NCAM）并加入Head模块中。通过与整体网络一起训练，得到归一化（batch normalization，BN）缩放因子，利用缩放因子算出各个通道的权重因子，提升网络对交通标志特征的表达能力，从而使YOLOv7-NCAM网络模型能够集中关注检测目标交通标志。通过在CCTSDB-2021交通标志检测数据集上的测试，与YOLOv7网络模型对比结果表明，YOLOv7-NCAM算法对背景复杂下小交通标志的检测各项指标均有明显提高：准确率（precision，P）达到91.5%，比原网络高出9.5个百分点；召回率（recall，R）达到85.9%，比原网络高出5.7个百分点；均值平均精度（mean average precision，mAP）达到了91.4%，比原网络高出4.7个百分点。与现有的交通标志检测算法相比，YOLOv7-NCAM算法的检测准确率也有提高，且检测速度48.3 FPS，能满足实时需求。

关键词: YOLOv7, 归一化通道注意力机制, 交通标志, 激活函数

Abstract: Aiming at the problem that the existing target detection algorithms are not effective in detecting small traffic signsin complex background, a YOLOv7 based on normalized channel attention mechanism (YOLOV7-NCAM) for detecting traffic signs is proposed. The YOLOv7-NCAM algorithm first constructs two types of convolutional layers, which are called DBF and CBF, using the FReLU activation function, and then uses them to form the Backbone module and the Neck module of the model. Then, a normalized channel attention mechanism (NCAM) is proposed and added to the Head module. By using the attention operator and the normalized (BN) scale factor, the NCAM mechanism can reduce the weight of the complex background to weaken the interference of complex background information and enhance the representation ability of target features, which allows the network to pay more attention to the traffic signs in the image. The test results on the CCTSDB 2021 dataset show that YOLOv7-NCAM achieves advantages over the YOLOv7 network in terms of various indexes： the accuracy (P) reaches 91.5%, which is 9.5 percentage points greater than that of the YOLOv7; Recall rate (R) reaches 85.9%, being 5.7 percentage points greater than that of the YOLOv7; and the mean average precision (mAP) reaches 91.4%, exceeding that of the YOLOv7 by about 4.7 percentage points. Furthermore, YOLOv7-NCAM achieves the accuracy results significantly superior to those of the state-of-the-art algorithms in detecting small traffic signs in complex background and meets real-time requirements with the detection speed of 48.3 FPS.

Key words: YOLOv7, normalized channel attention mechanism (NCAM), traffic sign, activation function

刘晶, 刘俊伟. 应用归一化通道注意力机制的YOLOv7交通标志检测算法[J]. 计算机工程与应用, 2025, 61(11): 249-258.

LIU Jing, LIU Junwei. YOLOv7 Traffic Sign Detection Algorithm with Normalized Channel Attention Mechanism[J]. Computer Engineering and Applications, 2025, 61(11): 249-258.

参考文献

[1] SHAOQING R, KAIMING H, ROSS G, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[2] REDMON J, DIVVALA K S, GIRSHICK B R, et al. You only look once: unified, real-time object detection[J]. J-STAGE, 2016, 18: 779-788.
[3] LIU W, ANGUELOV D , ERHAN D, et al. SSD: single shot multibox detector[J]. J-STAGE, 2016, 18: 21-37.
[4] 梁鑫, 代倩.基于注意力机制的改进YOLOv3与交通标志检测[J].现代计算机, 2022, 28(15): 9-16.
LIANG X, DAI Q. Improved YOLOv3 based on attention mechanism and its application in traffic sign detection[J]. Modern Computer, 2022, 28(15): 9-16.
[5] YAO Y B, LI H, DU C J, et al.Traffic sign detection algorithm based on improved YOLOv4-tiny[J]. Signal Processing: Image Communication, 2022, 107: 116783.
[6] 李杰, 高尚兵, 胡序洋, 等.基于DL-SSD模型的交通标志检测算法[J]. 扬州大学学报(自然科学版), 2022, 25(5): 47-53.
LI J, GAO S B, HU X Y, et al. Traffic sign detection algorithm based on DL-SSD model[J]. Journal of Yangzhou University (Natural Science Edition), 2022, 25(5): 47-53.
[7] 郎斌柯, 吕斌, 吴建清, 等.基于CA-BIFPN的交通标志检测模型[J].深圳大学学报(理工版), 2023, 40(3): 335-343.
LANG B K, LU B, WU J Q, et al. A traffic sign detection model based on CA-BIFPN[J]. Journal of Shenzhen University (Science And Engineering), 2023, 40(3): 335-343.
[8] 石镇岳, 侯婷, 苏勇东.改进YOLOv7的交通标志检测算法[J].计算机系统应用, 2023, 32(10): 157-165.
SHI Z Y, HOU T, SU Y D. Improved YOLOv7 traffic sign detection algorithm[J]. Computer Systems & Applications, 2023, 32(10): 157-165.
[9] REDMON J , FARHADI A. YOLO9000: better, faster, stronger[J]. J-STAGE, 2017, 19: 6517-6525.
[10] WANG Z, ZHU H, JIA X, et al. Surface defect detection with modified real-time detector YOLOv3[J]. Journal of Sensors, 2022, 2022: 8668149.
[11] ZHAO J, TIAN G, QIU C, et al.Weed detection in potato fields based on improved YOLOv4: optimal speed and accuracy of weed detection in potato fields[J]. Electronics, 2022, 11(22): 3709.
[12] ZHANG R, ZHENG K, SHI P, et al. Traffic sign detection based on the improved YOLOv5[J]. Applied Sciences, 2023, 13(17): 9748.
[13] YE G, QU J, TAO J, et al. Autonomous surface crack identification of concrete structures based on the YOLOv7 algorithm[J]. Journal of Building Engineering, 2023, 73(15): 106688.
[14] 成浪, 敬超.基于改进YOLOv7的X线图像旋转目标检测[J].图学学报, 2023, 44(2): 324-334.
CHENG L, JING C. Rotating target detection in X-ray images based on improved YOLOv7[J]. Journal of Graphics, 2023, 44(2): 324-334.
[15] OUYANG W, ZENG X, WANG X, et al. DeepID-Net: deformable deep convolutional neural networks for object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(7): 1320-1334.
[16] VARSHNEY M, SINGH P. Optimizing nonlinear activation function for convolutional neural networks[J]. Signal, Image and Video Processing, 2021, 15(6): 1-8.
[17] SZEGEDY C, LIU W, JIA Y, et al. Going deeper with convolutions[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, 17: 1-9.
[18] HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[J]. J-STAGE, 2015, 17: 770-778.
[19] HUANG G, LIU Z, LAURENS V D M , et al.Densely connected convolutional networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 2261-2269.
[20] HE K M, ZHANG X Y, REN S Q, et al.Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9): 1904-1916.
[21] GUO L, HE J, LIN P, et al.TRScore: a three-dimensional RepVGG-based scoring method for ranking protein docking models[J]. Bioinformatics, 2022, 38(9): 2444-2451.
[22] IOFFE S, SZEGEDY C.Batch normalization: accelerating deep network training by reducing internal covariate shift[J]. arXiv:1502.03167, 2015.
[23] 苏树智, 蒋博文, 陈润斌.基于实例归一化的通道注意力模块[J].计算机仿真, 2024, 41(1): 227-231.
SU S Z, JIANG B W, CHEN R B.Channel attention module based on instance normalization[J]. Computer Simulation, 2024, 41(1): 227-231.
[24] 尹宋麟, 谭飞, 周晴, 等.基于改进YOLOv4模型的交通标志检测[J].无线电工程, 2022, 52(11): 2087-2093.
YIN S L, TAN F, ZHOU Q, et al. Traffic sign detection based on improved YOLOv4 model[J]. Radio Engineering, 2022, 52(11): 2087-2093.
[25] 张达为, 刘绪崇, 周维, 等.基于改进YOLOv3的实时交通标志检测算法[J].计算机应用, 2022, 42(7): 2219-2226.
ZHANG D W, LIU X C, ZHOU W, et al. Real-time traffic sign detection algorithm based on improved YOLOv3[J]. Journal of Computer Applications, 2022, 42(7): 2219-2226.
[26] 李禹纬, 付锐, 刘帆.改进YOLOv7的轻量化交通标志检测算法[J].太原理工大学学报, 2024, 55(1): 195-203.
LI Y W, FU R, LIU F. Traffic sign detection algorithm based on improved YOLOv7[J]. Journal of Taiyuan University of Technology, 2024, 55(1): 195-203.
[27] 杨祥, 王华彬, 董明刚.改进YOLOv5的交通标志检测算法[J].计算机工程与应用, 2023, 59(13): 194-204.
YANG X, WANG H B, DONG M G. Improving the trafficsign detection and calculation method of YOLOv5[J]. Computer Engineering and Applications, 2023, 59(13): 194-204.