| With the rise of intelligent transportation field,traffic sign recognition as its key technology,the development is also rapid.However,the traditional traffic sign recognition requires a large amount of supervision data,which is difficult to obtain.Moreover,the traditional recognition method can only identify the new traffic sign under the high cost of retraining.So it cannot adapt to the regional traffic sign differences and meet the future needs.In fact,the traffic sign with a prototype image,so we use the method of few shot learning to solve the above problem.The current variable prototype encoder(VPE),which uses few shot learning in combination with the encoder to induce real and prototype images in the same metric domain,learns the image similarity and applies it to new classes to effectively deal with the problem.VPE uses a variational auto-coding structure to implicitly introduce a latent feature space,in which features from real data form a tight cluster around the feature points of the corresponding prototype,while VPE has obvious defects and improvements in spatial tight clusters and separability space.The main improvement content of this paper is three parts:first,in the data preprocessing stage,we use automatic data enhancement to find the best image transformation strategy,so as to improve the performance of the model without expanding the data set or generating new data.In addition,The data enhancement strategy he learned can be extended on the same type of data set to achieve better results;secondly,the idea of GAN is added to the VPE to make its reconstructed image more similar to the prototype image,and to strengthen its encoder and decoder;finally,in the model loss part,the KL divergence constraint is used to make the distribution of the real image and the prototype image in the hidden layer closer,improve the similarity within the class,and make the traffic sign image more separable.The few shot traffic sign recognition based on VPE model has achieved remarkable results in recognition and generation of traffic signs,which fully proves its superiority. |