Learning relations in human-like style for few-shot fine-grained image classification

Fine-grained classification is a challenging problem with small inter-class variance and large intra-class variance. It becomes more difficult when only a few labeled training samples are available. Inspired by the procedure of human recognition that two similar objects are usually distinguished by comparing their key parts, we develop a novel few-shot fine-grained classification method, which learns to model the inter-class boundaries in human-like style, i.e., extracting key-part structure information of objects and performing part-by-part comparison. To this end, we first extract the key parts of objects by using the designed key-part detector, which are then encoded by our structure encoder for the final comparison. To tackle with the scarce labeled samples, we train the proposed network under the metric-based few-shot learning methodology. Experiments on benchmark datasets demonstrate the effectiveness of the proposed method compared with the state-of-the-art counterparts. Besides, extensive investigations are conducted to verify the contributions of the key components of our method.