Background-aware decoupling and multi-granularity fusion-reconstruction approach for infrared small target detection.

Abstract

In infrared small target detection (IRSTD) tasks, small targets are challenging to distinguish due to their minute size, sparse pixels, and features that closely resemble complex background textures. Existing detection methods predominantly employ multi-layer downsampling to extract small target features directly from infrared images. However, the merging and discarding of pixels during downsampling reduces the resolution of feature maps, causing small target pixels to be smoothed, obscured, or even lost. This leads to false alarms and missed detections of small targets. To address these issues, this paper proposes a novel background-aware decoupling and multi-granularity fusion-reconstruction approach for infrared small target detection (BAMFNet). To address the difficulty in distinguishing complex background features from small target features, a decoupling algorithm separates infrared images into background and small target components. For extracting complex features from the background component, the Global-Local hierarchical fusion and Patch-Aware Attention Mechanism (GL-PAAM) utilizes context dependencies to adaptively focus on critical background regions. This refines and enhances features in the decoupled background component, reducing background interference in small target features. To preserve pixel-level details in small target components, the dilated convolution mechanism captures edge details without resolution loss, extracting sparse small target features. This mitigates the smoothing, masking, or loss of small target pixels during downsampling. Finally, multi-granularity fusion is performed on the extracted background and small target features to reconstruct a feature map with high contrast between small targets and background, further improving detection accuracy. Extensive comparative experiments against existing state-of-the-art infrared small target detection methods on three public datasets validate the effectiveness of the proposed method.