Abstract
Detecting infrared dim small targets remains a challenging task due to strong background interference and low signal-to-clutter ratio (SCR), which frequently results in target features being submerged by clutter. Furthermore, existing methods often struggle to strike a balance between physical interpretability and advanced feature representation. To address these limitations, this paper proposes a Deep Unfolding Residual Decomposition (DURD) method based on a novel dual-branch framework. First, an adaptive structured decomposition (ASD) module is proposed, which unfolds iterative optimization into a deep neural network. The ASD module leverages an implicit low-rank subspace projection and an adaptive structured sparse perception mechanism to separate high-quality physical sparse priors from low-rank background components. Second, to further exploit deep semantic representations and suppress high-frequency clutter, a residual-driven multi-scale Encoder–Decoder (RMED) is designed. Using physically constrained residual features as input, RMED emphasizes target regions while suppressing background interference. Through full-scale feature aggregation and deep supervision mechanisms, it hierarchically extracts robust multi-scale semantic representations. Finally, a dual-branch cooperative fusion (DCF) module is developed to integrate the outputs of both branches. The DCF employs a learnable gating factor to adaptively balance the contributions of physical sparse priors and deep semantic features, thereby achieving a complementary and discriminative enhancement for infrared small target detection. Extensive experiments demonstrate that the proposed DURD significantly enhances the detection accuracy in complex scenarios. The code is available at https://github.com/haofan72/DURD.
Funding Information
  • Fundamental Research Funds for the Central Universities (510224078)

This publication has 55 references indexed in Scilit: