Bi-channel attention meta learning for few-shot fine-grained image recognition

Abstract

Few-shot fine-grained recognition is an attractive research topic that aims to differentiate between sub-categories using a limited number of labeled examples. Due to the characteristics of fine-grained images, capturing subtle differences between categories using limited samples is very challenging. Discriminative information is essential for fine-grained image recognition, however, existing methods of few-shot learning usually extract features from each part indiscriminately, resulting in poor performance. To solve this problem, this work presents a compact Bi-channel Attention Meta-learning Model with an embedding module and a feature calibration module. The embedding module can effectively prevent the loss of crucial spatial information, thereby learning better deep descriptors. The feature calibration module consists of two sequentially arranged channel attention blocks, which allow the network selectively enhances discriminative features and compress less useful features with global information. Experiments on three commonly used fine-grained benchmark datasets indicate the efficacy and superiority of the proposed model.