Learning to focus: cascaded feature matching network for few-shot image recognition

Generally, deep networks learn to recognize a category of objects by training on large number annotated images accurately. However, meta-learning problem known as low-shot image recognition task occurs when few with annotations are available for learning model single category. Consequently, the in testing/query and training/support datasets likely vary terms size, location, style, so on. In this paper, we propose method, cascaded feature matching network (CFMN), solve problem. We train meta-learner more fine-grained adaptive distance metric using block, which aligns associated features together naturally ignores non-discriminative features. By applying proposed block different layers network, multi-scale information among compared is incorporated into final feature, boosts performance generalizes better relationships. Moreover, experiments few-shot (FSL) two standard datasets: miniImageNet Omniglot, confirm effectiveness our method. Besides, multi-label first studied new data split COCO dataset, further shows superiority performing FSL complex images.