Research on vehicle head-tail recognition based on DSDCVGG19 algorithm

With the development and implementation of convolutional neural networks in pattern recognition, there are large number of parameters needs to calculate and storage, which makes the algorithm hard to run in common computer. For the mine vehicle refinement recognition project we are currently studying, how to obtain and analyze the feature information of the target in the image on top of the limited computer arithmetic power, and maximize the accuracy and efficiency of the experimental research become the main purpose of the research. In this paper we propose a dense connectivity and deeply separable  VGG19  network model with three tricks to overcome this problem. The three tricks are convolution instead of full connectivity; the deeply separable convolutions; dense connectivity network module. The first two methods could reduce the parameters and calculate. The third trick makes the model have more features in limited parameters. The combination of the three improved methods achieves a better result in vehicle head and tail recognition. The experimental results show that the optimized model improves the recognition accuracy (4.41% compared to VGG19) while drastically reducing the parameters (71% compared to VGG19).