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

Automated summary

The paper proposes a cascaded feature matching network (CFMN) to tackle the low-shot image recognition task, training a more fine-grained and adaptive deep distance metric using feature matching block and incorporating multi-scale information among compared images in different layers to boost recognition performance and improve generalization. Experiments on few-shot learning using standard datasets confirm the effectiveness of the proposed method, and the study of multi-label few-shot task on a new data split of the COCO dataset further demonstrates the superiority of the feature matching network in complex images.