DUSRNet: Deep Unfolding Sparse-Regularized Network for Infrared Small Target Detection

Abstract

In the field of Infrared Small Target Detection (ISTD), deep unfolding-based techniques have demonstrated significant efficacy. However, existing methods that utilize low-rank sparse models for ISTD task tend to heavily emphasize the global low-rank characteristic of infrared image and ignore local structural feature. To solve the challenges of complex background and low signal-noise ratio, we propose a Deep Unfolding Sparse-Regularized Network termed as DUSRNet. It intuitively, combines the powerful feature extraction capability of deep learning with the fine structure description capability of sparse regularization over infrared image background. To adaptively describe the low-rank and sparse characteristic between background and target, the sparse-regularized infrared small target detection model is seamlessly embedded into the deep neural network in a end-to-end manner. We customize adaptive background estimation module, sparse target extraction module and infrared image reconstruction module to unfold the proposed model. Extensive experimental results demonstrate that our DUSRNet achieves state-of-the-art (SOTA) results on the public NUDT-SIRST, SIRST-Aug and IRSTD-1k datasets. Especially, compared with RPCANet,which also adopted deep unfolding method, on IRSTD-1k dataset with extremely high scene complexity and variability, the proposed method has an increase of 35.63%,19.05%,7.23% and $-$73.56% in mIoU, F1, Pd and Fa indexes, respectively.