DDFet: Infrared Small Target Detection via a Dual-Domain Fused Deep Unfolding Network

Infrared small target detection has a wide range of applications, such as target warning and maritime rescue. However, existing deep learning methods still encounter some limitations, including the effective utilization of domain priors and the generalization ability of networks. To address these issues, we propose a dual-domain fused deep unfolding network (DDFNet) based on half-quadratic splitting (HQS) optimization, which offers better inner mechanism. First, we developed a network-learned implicit regularization method for infrared background to mitigate the issue of insufficient prior knowledge acquisition. Subsequently, the optimization algorithm is unfolded into a learnable deep network, with each module representing a solving operator during the iterative process, enabling automatic updates and solutions within the network. Specifically, we integrate spatial and frequency domain prior information to enhance the accuracy of background estimation. Additionally, we design a hyperparameter module to adaptively learn all regularization parameters through the network, addressing the difficulty of manual parameter tuning in optimization methods. In this manner, our network inherits the powerful learning capabilities of deep learning methods while retaining the prior information and interpretable mechanisms of traditional optimization methods. Experimental results demonstrate that our proposed method exhibits significant advantages in both performance and generalization ability compared to existing state-of-the-art approaches. Our code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/AuroraPei/DDFNet</uri>.