Multifarious support-and-query feature interaction network for few-shot anomaly segmentation

Timely anomaly detection is crucial in industrial settings, with numerous detection methodologies relying heavily on extensive data collections. Nonetheless, the scarcity of data poses a significant challenge when encountering novel anomaly types, especially when only a limited number of instances are available for analysis. To address this, we introduce a Multifarious Support-and-Query Feature Interaction Network (MSFINet), the few-shot learning-based approach for identifying anomalies in key parts of high-speed trains. Specifically, MSFINet introduces a Dual-attention Feature Amplifier (DFA), which strengthens varied feature hierarchies by integrating self-attention and mutual attention mechanisms. The Mask-based Commonality Convergence (MCC) is designed to integrate contextual understanding, pinpoint the similarity of support images and query images, and converge the features with support mask, thereby enhancing discriminative power. Lastly, the Comprehensive Correlation Learner (CCL) is incorporated to learn a series of correlation maps that encapsulate both high-level semantic structures and fine-grained, low-level details. Our approach’s efficacy is substantiated by its superior performance on our tailored anomalous dataset and the PASCAL- 5 i benchmark dataset. Specifically, our technique outpaces the leading-edge solutions by a considerable extent on the anomalous dataset.