Multi-Depth Learning with Multi-Attention for fine-grained image classification

Abstract

Compared with the traditional image classification task, fine-grained image classification has the difficulty of small differences between classes and large differences within classes. In view of this difficulty, attention proposal has been widely used in fine-grained image classification. However, traditional attention proposal has to localize first and then processing. Model needs to run step by step and the attention focusing method is single. This paper proposed a model (MAMDL, Multi-Attention-Multi-Depth-Learning) which combines multiple attention mechanisms and multi network parallel learning. The advantage of MAMDL is that it can first learn end-to-end. Secondly, the multiple attention mechanisms can effectively combine four attention mechanisms to improve the network’s ability to process local features. Finally, this paper focuses on the attention found in the backbone network, Feature extraction from branch convolution neural networks with different depths enhances the classification performance of the model. The experimental results show that MAMDL outperforms mainstream fine-grained image classification methods on the fine-grained image classification dataset CUB-200, Stanford dogs and Stanford cars.