Attention-based supervised contrastive learning on fine-grained image classification

Abstract

To solve the problem of fine-grained image classification performance caused by intra-class diversity and inter-class similarity in fine-grained images, we propose an Attention-based Supervised Contrastive (ASC) algorithm for fine-grained image classification. The method involves three stages: firstly, local parts are generated by a multi-attention module for constructing contrastive objectives to filter useless background information; an attention-based supervised contrastive framework is introduced to pre-train an encoder network and learn generalized features by pulling positive pairs closer while pushing negatives apart. Finally, we use cross-entropy to fine-tune the model pre-trained in the second stage to obtain classification results. Comprehensive experiments on CUB-200-2011, FGVC-Aircraft, and Stanford Cars datasets demonstrate the effectiveness of the proposed method.