BDSFusion: A bidirectionally driven saliency fusion network for enhancing small target detection in infrared dual-band images

Single-band infrared images provide limited scene information, whereas dual-band and multiband infrared images offer more comprehensive and complementary scene data, enhancing the accuracy and robustness of infrared small target detection (IRSTD). To address the challenges of medium-wave infrared images suffering from substantial high-intensity background clutter and long-wave infrared images exhibiting a low signal-to-clutter ratio (SCR) for small targets, we proposed a bidirectionally driven saliency fusion network, termed BDSFusion. Specifically, to improve the local feature extraction capability and reduce the channel redundancy of the vanilla Mamba module, we proposed a feature extraction module called CSA-CM, which consists of a Conv-Mamba (CM) module and a Channel Self-attention (CSA) module. To enhance the contrast between target and background, we proposed a spatial-frequency cross-domain feature integration module (SFCIM), which leverages the complementary characteristics of spatial and frequency-domain information from source images. Additionally, to address the lack of task-specific loss functions in IRSTD, a task-driven joint loss function that combines intensity loss, Haar wavelet mask loss, and background residual loss, to guide the fusion network in suppressing background clutter and enhancing target information. Extensive fusion experiments on our custom dataset demonstrated that our method outperformed existing fusion methods in both qualitative and quantitative evaluations, leading to improved IRSTD performance. To further validate the generalization capability of our approach, we extended and evaluated BDSFusion on the task of infrared-visible image fusion for small target detection. The method consistently demonstrated strong performance, confirming its robust and adaptable design. The code is available at https://github.com/kyrietop11/BDSFusion.