一种Yolov5颈部细化的小交通标志检测算法

A small traffic sign detection algorithm for improved Yolov5 neck refinement

针对现有交通标志检测算法对小尺寸交通标志特征提取不充分、检测精度低、速度慢等问题,文中提出一种Yolov5颈部细化的小交通标志检测算法。首先,为使网络更关注小交通标志的检测,在主干网络中添加一条浅层特征提取分支,获得浅层特征图,并与具有较强语义信息的深层特征图融合,改善浅层特征图的感受野;其次,在Yolov5颈部网络引入GSConv模块,通过深度可分离卷积与普通卷积信息渗透、网络通道减半压缩的方式提高算法的检测性能;最后,利用完备交并比(CIoU)损失函数加快模型收敛,提高检测速度。实验结果表明,改进后的Yolov5网络模型性能有所上升,在TT100K交通标志数据集上平均精度均值mAP可达83.56%,相较于原始Yolov5基本框架mAP提升2.24%,且检测速度FPS可达40.5 f/s,满足实时性要求。

In allusion to the problems of insufficient feature extraction,low detection accuracy and slow speed of the existing traffic sign detection algorithm for small traffic signs,a small traffic sign detection algorithm with Yolov5 neck refinement is proposed.In order to make the network pay more attention to the detection of small traffic signs,a shallow feature extraction branch is added to the backbone network to obtain the shallow feature map,and integrated with the deep feature map with strong semantic information to improve the feeling field of the shallow feature map.The GSConv module is introduced into the Yolov5 neck network to improve the algorithm′s detection performance by means of deep separable convolution and ordinary convolution information penetration,and network channel half compression.The complete intersection over union(CIoU)loss function is used to accelerate the model convergence and improve the detection speed.The experimental results show that the performance of the improved Yolov5 network model has improved,the mean value mAP of average accuracy on the TT100K traffic sign dataset can reach 83.56%,which is 2.24%higher than the mAP of the original Yolov5 basic framework,and the detection speed can reach 40.5 f/s,meeting the real-time requirements.