Abstract: Finding discriminative local regions corresponding to fine-grained features is the key to solve fine-grained image classification problems. Recently， fine-grained classification by weakly supervised data augmentation network（WS-DAN）has achieved excellent results，but its single cross-entropy loss（CE-Loss）makes the network focus on global discriminative regions and misses some local discriminative regions. To address this problem，this paper proposed a data augmentation network （MC-DAN） based on mutual channel loss（MC-Loss），which could force feature channels belonging to the same class to be more discriminative. Second，a counterfactual attention mechanism（CAL）was introduced to encourage the network to learn more attention information by counterfactual intervention. In addition， an attention module combining spatial attention and channel attention was proposed to better focus on object regions in images. Comprehensive experiments on three public datasets showed that the method could effectively achieve classification.

Key words: fine-grained image classification, mutual channel loss, counterfactual attention learning, data augmentation