Sparse Prior Is Not All You Need: When Differential Directionality Meets Saliency Coherence for Infrared Small Target Detection

TL;DR

This paper tackles infrared small target detection in cluttered scenes, where traditional tensor methods can struggle to separate targets from background. It introduces Sparse Differential Directionality (SDD), which leverages directional priors on background versus non-directional targets, and couples this with an Adaptive Saliency Coherence (ASCE) map to boost target contrast. The background is modeled via a Tucker-based low-rank factorization with mixed directional sparsity, while the sparse target component is constrained by saliency-informed weights; a Proximal Alternating Minimization (PAM) solver guarantees convergence in a non-convex setting. Extensive experiments on real infrared sequences show the method consistently outperforms ten state-of-the-art approaches in detection and clutter suppression, with code available at https://github.com/GrokCV/SDD, highlighting its practical impact for robust IRSTD systems.

Abstract

Infrared small target detection is crucial for the efficacy of infrared search and tracking systems. Current tensor decomposition methods emphasize representing small targets with sparsity but struggle to separate targets from complex backgrounds due to insufficient use of intrinsic directional information and reduced target visibility during decomposition. To address these challenges, this study introduces a Sparse Differential Directionality prior (SDD) framework. SDD leverages the distinct directional characteristics of targets to differentiate them from the background, applying mixed sparse constraints on the differential directional images and continuity difference matrix of the temporal component, both derived from Tucker decomposition. We further enhance target detectability with a saliency coherence strategy that intensifies target contrast against the background during hierarchical decomposition. A Proximal Alternating Minimization-based (PAM) algorithm efficiently solves our proposed model. Experimental results on several real-world datasets validate our method's effectiveness, outperforming ten state-of-the-art methods in target detection and clutter suppression. Our code is available at https://github.com/GrokCV/SDD.

Figures (18)

• Figure 1: Illustration of directional information of background and target in infrared sequences. (a) Infrared Sequence. (b) Unfolding the four scenes in the sequence horizontally. (c) Difference between adjacent pixels during horizontal unfolding. (d) Unfolding the four scenes in the sequence vertically. (e) Difference between adjacent pixels during vertical unfolding.
• Figure 2: Illustration of 3-D tensor Tucker decomposition.
• Figure 3: The difference images along the temporal mode for infrared sequence data in both spatial modes. (a) The heatmaps of the spatial horizontal difference, (b) The heatmaps of the spatial vertical difference, (c) and (d) The matrices of corresponding difference images along the temporal mode, (e) and (f) Display of the shared and unshared sparsity.
• Figure 4: Illustration of single and multiscale saliency coherence maps. (a) Original image (b)-(d) The saliency map obtained by the first to three saliency coherence items. (e) The saliency map obtained by the transformed enhancement factor.
• Figure 5: The overall flow-process diagram of the proposed infrared small target detection model. During detection, the target saliency enhancement factor ${{\cal W}_{ASCE}}$, obtained by stacking single frame ASCE mapping, is first constructed and then integrated into the SDD model based on PAM optimization. During decomposition, the target components from each layer are multiplied by the enhancement factor to highlight targets and suppress background clutter.