Global and Local Attention Embedding Network for Few-Shot Fine-Grained Image Classification

Abstract

Few-shot fine-grained image recognition aims to classify fine-grained images with limited training samples. Nowadays exist in a majority of few-shot fine-grained image classification methods the following problems: local information loss and ignoring pivotal parts. To solve the above problems, this paper proposes a new embedding module, called GLAE. The author designs a hierarchical structure and combines the first-order and second-order information to reduce the local information loss. Besides, this paper proposes an attention mechanism to obtain the vital parts by the attention mask. On the StanfordCars dataset, GLAE achieves an accuracy of 91.18% which is the best result in the field of few-shot fine-grained image recognition.