A Progressive Multi-Scale Relation Network for Few-Shot Image Classification

Abstract

Few-shot classification addresses the challenge of swiftly enabling a deep learning model to comprehend new classes based on minimal supporting image samples. Despite recent research predominantly focusing on devising increasingly intricate classifiers for gauging similarities between query and support images, there is room for improvement in the critical role of feature embeddings and relational measurements. In this paper, a novel technique called Progressive Multi-Scale Relation Network (PMRNet), is designed for few-shot classification tasks. Our work introduces a feature extraction module that adeptly extracts and interactively fuses multi-scale features to address the problem of insufficient feature extraction. Furthermore, an enhanced attention mechanism is integrated, which amplifies critical local and global features for increased discriminative power. A novel bilinear relation module is also designed specifically to accurately quantify the similarities between the extracted features. The PMRNet has been rigorously evaluated on three benchmark datasets. PMRNet’s empirical performance in few-shot image classification tasks exhibits commendable accuracy, substantiating its efficacy as a robust architectural framework for significantly boosting the precision of such tasks.

Keywords

• Few-shot learning
• image classification
• metric learning
• multi-scales feature
• attention module