Prompt and Refinement: Asymmetric Mutual Learning for Infrared Small Target Detection with Noisy Labels
This study addresses the issue of noisy labels causing models to learn spurious patterns in infrared small target detection by proposing a robust training framework based on asymmetric mutual learning. The method introduces a novel asymmetric mutual correction mechanism between a task-specific detector and the Segment Anything Model (SAM) foundation model, effectively mitigating self-confirmation bias. By integrating prompt tuning, evidential uncertainty estimation, and local contrast regularization, the framework achieves high-quality label correction. Experimental results demonstrate that the proposed approach attains state-of-the-art performance across multiple noise scenarios on three benchmark datasets, significantly enhancing detection accuracy.