Abstract:

To solve the problem of the low recognition rate caused by the redundancy feature on the image classification, an algorithm for identification of fine-grained vehicles based on singular value decomposition and central metric is proposed. First, a convolutional neural network based on singular value decomposition is designed. With the method, the weight matrix of the fully connected layer is decomposed by the singular value, then it is re-assigned and fine-tuned. In this way, the redundant features with correlation can be removed, and the discriminative features of fine-grained levels can be learned. Second, a fusion loss method for different learning features is utilized, in which the losses of the central distance and the classification are fused in a weighted way to shorten the distance between the learned feature classes. Finally, experimental results prove that the test accuracy of the Residual Network (ResNet) framework is up to 93.02% in the Cars-196 fine-grained model data set. The proposed model outperforms the bilinear and attention model. Furthermore, the extended experiments prove that the method is applicable to other network frameworks.

Key words: fine-grained vehicle recognition, redundancy feature, singular value decomposition convolutional neural networks, fusion loss