Abstract: Fine-grained image classification is very challenging due to the limited amount of data with large intra-class differences and small inter-class differences. Since the deep features have strong feature representation capability and the middle layer features can effectively supplement the missing information of the global-level features in fine-grained image identification, in order to take advantages of the convolutional layer feature, this paper proposes a channel adaptive discriminative learning method for fine-grained image classification. In this method, the intermediate features are first aggregated in the channel direction to capture the target position, and then we classify the information obtained by the interactive cascade of the region of interest features. Finally, the proposed method can perform end-to-end training without any bounding box and part annotation. A large number of experiments on three common fine-grained image classification datasets (CUB-200-2011, Stanford Cars and FGVC-Aircraft) have shown that this method can not only maintain simple and reasonable efficiency, but also improve the accuracy, compared with the other methods.

Key words: fine-grained image classification, channel adaptive, mask, feature enhancement, interested area