Decoupling polarity with self-prompts: A new framework for bright–dark infrared small target detection

Detecting dim and small infrared targets in complex, dynamic backgrounds is challenging because of grayscale inversion and target feature heterogeneity. To address this, we introduce SPBD-Det, a novel detection framework inspired by visual multi-pathway principles. SPBD-Det first employs a steady-state partial differential equation (PDE) model that combines Sobel-based edge enhancement with an implicit noise learning module. This approach prevents information loss during downsampling and significantly improves the edge contrast of small targets. Furthermore, we propose a Self-Prompt Decoder (SPD) that uses learnable tokens and cosine distance constraints to adaptively decouple bright and dark target features, enabling targeted feature selection for improved accuracy. Extensive evaluations on two public benchmarks (IRTiny-BD-10K and SIRST-Aug) and on our constructed IRReversal dataset demonstrate that SPBD-Det achieves superior performance and significantly outperforms state-of-the-art methods in detection accuracy and segmentation quality, while exhibiting strong robustness across varying scene complexities. Code and dataset are available at https://github.com/Linaom1214/SPBD-Det.