Learning Multiple Receptive Fields for Few-Shot Image Classification

The size of the receptive field of current backbone feature extractor networks (e.g., Conv4Net, Conv64F and ResNet12) used in existing few-shot learning methods is fixed and lacks adaptation to changes in the scale of the target object, leading to poor performance when measuring the similarity of objects with different scales of the same category. In this paper, a novel parallel architecture, Feature United Neural Network (FUNNet), is proposed to learn the features with multi-receptive fields for visual representation. The proposed network is constructed by parallelly combining multiple branches with different receptive field sizes. To this end, we propose k-Conv-Block and Identity Block as the branch component. Each branch is constructed by cascading multiple k -Conv block and identity block. To identify the local feature of the target object, we propose a similar feature selection module. The decision is made by comparing the similarity between feature matrices from the support and query samples. Extensive experiments on five publicly available benchmark datasets consistently demonstrate that the proposed FUNNet achieved significant increases in accuracy over the latest state-of-the-art methods in few-shot image classification.