基于MobileNetV3-YOLOv4的车型识别

Vehicle Type Recognition Based on MobileNet-YOLOv4

为改善现有车型识别算法在参数量较大时训练时间较长等缺点,提出基于原有YOLOv4的改进算法。通过使用MobileNetV3替换YOLOv4的主干特征提取网络,使用深度可分离卷积替换标准卷积的方式削减模型参数量,然后使用K-means算法设置预选框用以提升模型最终识别精度,其中MobileNetV3部分在训练模型时使用迁移学习的方法,加快了整个模型的收敛速度。实验结果表明,改进算法在BIT-Vehicle数据集上的车型识别准确率为96.17%,参数量约为53.77MB,检测速度较YOLOv4提升了26%。改进识别算法在保证95%精确度的情况下,降低了车辆识别模型的参数量,同时也提升了检测速度。

In order to improve the shortcomings of existing vehicle recognition algorithm such as longer training time due to numerous parameters,proposes an improved algorithm based on the original YOLOv4.By using MobileNetV3 to replace the backbone feature ex⁃traction network of YOLOv4,and adopting deep separable convolution instead of standard convolution to reduce the number of parame⁃ters in the model,and finally the K-means algorithm is used to set the preselection box to improve the accuracy of the model.In partic⁃ular,the transfer learning method is used to train the model in the MobileNetV3 part,which accelerates the convergence speed of the entire model.The experimental results show that the accuracy of the vehicle identification method in this paper on the BIT-Vehicle da⁃taset is 96.17%,the parameter amount is about 53.77MB,and the detection speed is 26%higher than the basic YOLOv4.The im⁃proved vehicle recognition algorithm reduces the number of parameters of the vehicle recognition model and greatly increases the detec⁃tion speed in the case of guaranting 95%accuracy.