Star-Graph Interaction Network for Multi-Input Infrared Small Target Detection

Abstract

In infrared small target detection (IRSTD), data-driven methods typically utilize a single-scale image as input, complicating the suppression of background clutter during early encoding. While multi-input methods alleviate this issue, their independent encoding of inputs often fails to suppress undesirable representations at each scale, reducing segmentation stability. Furthermore, high-level representations lack background context after multiple downsamplings, leading to overactivation of background clutter in the decoding stage. To address these challenges, we propose a star-graph interaction network (SGINet), an alternative paradigm for multi-input IRSTD. Specifically, all inputs are encoded as satellite nodes, and a residual attention convolution module masks noisy channels and spatial regions, enhancing target–background contrast during early encoding. The star-graph interaction independently updates the satellite nodes, effectively suppressing undesirable representations at each scale. In addition, we introduce a background complement module to supplement the lacked background context of high-level hub representation, preventing the overactivation of background clutter in the output hub representation. SGINet outperforms recent methods on three datasets, NUAA-SIRST, NUDT-SIRST, and IRSTD-1k, especially in segmentation accuracy and inference speed.