Fine-Grained Image Classification with Coarse and Fine Labels on One-Shot Learning

Abstract

In this paper, we aim to solve the fine-grained image classification on one-shot learning, which only has one image provided from each class. Specifically, we introduce the hierarchical structure between coarse and fine labels to exploit the relationship among categories. First, we make coarse label prediction of the input image and utilize Attention Proposal Network (APN) to determine the attentive area for fine label prediction. Then, according to the result of coarse label prediction, we can automatically select the images belong to the same coarse category from all samples in the support set to form a subset, which will be sent to relation network. Finally, we fuse the results of relation network and those of fine label prediction to produce more robust and more accurate classification results. The superior fine-grained classification performance of our method is demonstrated on CUB-200-2011 dataset and miniImageNet dataset.