Image Classification Algorithm combining Local Feature and Attention

In order to solve the problem of low accuracy of image classification caused by incomplete feature information extracted by convolutional neural network and indistinguishable feature information, an image classification algorithm combining local feature and attention was proposed. Firstly, patch layer is added to obtain multiple patch images before extracting image features, and feature extraction is carried out at both global and local levels. VGG16 is used as the backbone network to extract patch image features. Then, for each patch image, feature attention is carried out in the channel domain and spatial domain respectively to obtain important patch channel information and spatial information. In addition, in order to further constrain the feature attention of patch image, the patch attention aggregation loss is proposed to penalize the weights of irrelevant regions and strengthen the attention of important patch regions, which makes the extracted feature information more discriminative, so as to facilitate classification and improve the final classification performance. The experimental results on CIFAR-10, CIFAR-100 and RAF-DB datasets show that the classification performance of the proposed algorithm is significantly improved compared with other advanced methods.