Local Contrast Prior-Guided Cross Aggregation Model for Effective Infrared Small Target Detection

Infrared small target detection, referring to discovering the precise shapes of dim targets from complex clutter background, has gradually become a hot spot. In recent years, learning-based methods have become the mainstream schemes with high efficiency. However, these methods seldom consider the nature characteristics of infrared targets and are with high demands on computational complexity. By investigating the local contrast property, we propose the deformable attention module to separate salient and representative features and reduce the complexity of calculation. Furthermore, in order to aggregate the global and local correlations, we present the cross aggregation combining the transformer and convolution modules with the supervision of targets edges. Comprehensive experiments on the IRSTD-1k dataset demonstrate the superiority of our method, improve 3.32% of IOU and accelerate 50.59% of inference time compared with the most advanced method.