LCPNet: Latent Consistent Proximal Unfolding Network for Infrared Small Target Detection

LCPNet is a model-driven deep unfolding network for infrared small target detection. It moves the unfolding process into a latent representation, where the authors claim the low-rank prior remains valid while avoiding repeated reconstruction and compression of intermediate states. The method introduces a Latent Consistent Proximal solver with task-adaptive normalization and gain control, plus Shared Optimization Memory that provides a common historical state to decomposition variables. According to the paper, experiments on four public benchmarks show accurate and robust detection with low false-alarm rates and competitive efficiency among high-accuracy unfolding methods. Model and code are available on GitHub.