BPGA: Enhancing infrared small and dim target detection with background prior guide and vision graph attention

Infrared small and dim target detection is a crucial technology in various infrared search and warning systems, yet it faces significant challenges due to the targets’ small spatial scale and weak features, often obscured by complex backgrounds. Current approaches often struggle by focusing singularly on either target feature enhancement or background modeling, failing to synergistically leverage both aspects. To address this dual challenge, this paper proposes the BPGA (Background Prior Guide and Vision Graph Attention for Enhancing Infrared Small and Dim Target Detection) framework. BPGA innovatively introduces two key mechanisms: (1) Background prior guidance, leveraging the scene understanding capabilities of the Mobile SAM large visual model to suppress structured background clutter. (2) A vision graph convolution mechanism, realized through the proposed SA-Graph (Vision Graph Module based on Self-Attention) module, which adaptively constructs the image graph structure to obtain more stable and efficient feature information of small and dim targets, thereby amplifying subtle target signatures. Crucially, these components capture complementary prior information: Mobile SAM provides global scene context, while the SA-Graph module learns local contrast patterns. An attention fusion module then intelligently integrates this prior background information with the target feature information through self-attention, preventing semantic conflicts. Experimental results demonstrate that BPGA achieves State-of-the-Art (SOTA) performance on four open-source datasets: IRSTD-1k, Sirst AUG, MSISTD, and NUAA. The code is available at: https://github.com/Linaom1214/BPGA.