面向车型识别的YOLOv7网络优化研究

Optimization on YOLOv7 Network for Vehicle Type Identification

车型识别是智能交通系统的重要组成部分,针对YOLOv7网络应用于车型识别检测效果仍有提升空间,提出了一种改进的YOLOv7车型识别网络。首先,将主干网络的SiLU激活函数替换为漏斗激活函数F-ReLU,以扩大主干网络卷积层的感受野范围,使其在激活函数阶段实现空间信息建模;其次,在主干网络融入了多级注意力机制,以此加强主干网络对车型深层特征的提取能力及泛化能力;最后,将CIOU损失函数替换为Alpha-SIOU损失函数,以提升网络对车型目标的定位能力,并加速网络收敛。使用公开的BIT-Vehicle数据集进行测试,改进后的YOLOv7网络提升了车型识别性能,其中mAP0.5为98.4%,提升了0.8%;mAP0.5:0.95为96.1%,提升了1.2%;召回率为95.2%,提升了2.6%;检测速度为33.1帧/s,提升了1.9帧/s。

Vehicle type identification is an important part of intelligent transport systems.To improve the detection effect of vehicle type recognition by using YOLOv7 network,the improved YOLOv7 vehicle type recognition network was proposed.Firstly,the SiLU activation function of backbone network was replaced with the funnel activation function F-ReLU to expand the receptive field range of convolutional layers in the backbone network.It enabled spatial information to model in the activation function stage.Secondly,the multi-level attention mechanism was integrated into backbone network to enhance the extraction and generalization capabilities of deep features of vehicle types.Finally,the CIOU loss function was replaced with the Alpha-SIOU loss function to improve the network’s ability to locate vehicle targets and accelerate network convergence.The public BIT-Vehicle dataset was used for test.The improved YOLOv7 network improves vehicle type identification.The mAP0.5 is 98.4%with 0.8%increase.The value of mAP0.5:0.95 is 96.1%with 1.2%increase.The recall rate is 95.2%with 2.6%increase.The detection rate is 33.1 frames per second with the increase of 1.9 frames per second.