Hybrid Adjacency Representation for Few-Shot Classification

With the widespread use of large language models, image parsing remains a hot issue in the field. Although the deep learning method can solve some problems in an extent, specially domains with small dataset are hard to apply. This paper is dedicated to image classification using few-shot learning methods. First, it utilizes a basic framework for the entry classes of support and query sets, where the query sets are distinguished by the similarity of the class attribute. A hybrid adjacency representation(HAR) composed of a global self-similarity descriptor (GSSD) and a cross-similarity descriptor (CSD) is proposed. The GSSD is used to extract structural patterns from query and support sets, then transform the base representations into its self-correlation tensor. The CSD is adapted to aggregate the selected features of GSSD to promote the similarity of visual appearance. Lastly, the ablation studies and the extensive experiments show competitive results over the state-of-the-art methods on the benchmark.