Self-guided information for few-shot classification. (November 2022)

• Record Type: Journal Article
• Authors: Zhao, Zhineng; Liu, Qifan; Cao, Wenming; Lian, Deliang; He, Zhihai
• Is Part Of: Pattern recognition. Volume 131(2022)
• Publication Date: 2022-11
• Subjects: Few-shot classification -- Graph convolution network -- Self-guided information; 00-01 -- 99-00; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108880
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose the Self-Guided Information Convolution (SGI-Conv), which can accurately extract more distinctive features by regenerating high-level features mixture with low-level features. We design a hierarchical graph convolution network, graph convolution block network (GCBNet), sharing the adjacent matrix to deepen the depth of the graph convolutional network, which enhancing the aggregation ability. We conduct various experiments on the few-shot classification to demonstrate that our method achieves state-of-the-art performance on multiple benchmark datasets compared with other methods. Abstract: Few-shot classification aims to identify novel categories using only a few labeled samples. Generally, the metric-based few-shot classification methods compare the feature embedding of Query samples (unlabeled samples) with Support samples (labeled samples) in a metric algorithm to predict which category the Query sample belongs to. Obtaining a good feature embedding for each sample in the feature extraction stage can improve the classification accuracy in the metric stage. Based on this, we design the Self-Guided Information Convolution (SGI-Conv), an improved convolution structure, which utilizes the high-level features to guide the network to extract the required discriminative features. To effectively utilize the feature embeddings of samples, we divide the metric network into multiple blocks and build a multi-layer graph convolutional network by sharing adjacent matrices. The multi-layer structure enhances the aggregation ability of graph convolution. Extensive experiments on multiple benchmark datasets demonstrate that our method has achieved competitive results on the few-shot classification tasks.