CSPENet: Contour-Aware and Saliency Priors Embedding Network for Infrared Small Target Detection

TL;DR

CSPENet addresses infrared small target detection by explicitly embedding contour-aware and saliency priors into a deep framework. It introduces SCPEM to extract CP1 and CP2 priors, DBPEA to embed them in dual pathways, and AGFEM to refine multi-scale features, achieving state-of-the-art performance on three public ISTD datasets with improved $IoU$, $F_{1}$, $P_{d}$, and reduced $F_{a}$. The approach demonstrates robust localization and detailed contour preservation in dense clutter, while maintaining competitive parameter efficiency. This work offers a practical strategy for integrating structural priors into DL-based ISTD, potentially enabling more reliable detection in real-world infrared sensing scenarios.

Abstract

Infrared small target detection (ISTD) plays a critical role in a wide range of civilian and military applications. Existing methods suffer from deficiencies in the localization of dim targets and the perception of contour information under dense clutter environments, severely limiting their detection performance. To tackle these issues, we propose a contour-aware and saliency priors embedding network (CSPENet) for ISTD. We first design a surround-convergent prior extraction module (SCPEM) that effectively captures the intrinsic characteristic of target contour pixel gradients converging toward their center. This module concurrently extracts two collaborative priors: a boosted saliency prior for accurate target localization and multi-scale structural priors for comprehensively enriching contour detail representation. Building upon this, we propose a dual-branch priors embedding architecture (DBPEA) that establishes differentiated feature fusion pathways, embedding these two priors at optimal network positions to achieve performance enhancement. Finally, we develop an attention-guided feature enhancement module (AGFEM) to refine feature representations and improve saliency estimation accuracy. Experimental results on public datasets NUDT-SIRST, IRSTD-1k, and NUAA-SIRST demonstrate that our CSPENet outperforms other state-of-the-art methods in detection performance. The code is available at https://github.com/IDIP2025/CSPENet.

Figures (12)

• Figure 1: Visual comparison. Segmentation results of complex-structured targets (zoomed) in dense clutter environment obtained by different methods.
• Figure 2: Infrared small target image (upper) and its corresponding 3D local structural characterization visualization (lower).
• Figure 3: The overall architecture of the proposed CSPENet for infrared small target detection. 1) SCPEM: The input image is first fed into SCPEM to obtain two types of CP components. 2) DBPEA: The two types of CP components are embedded into the deep learning network through DBPEA, where CP1 is jointly encoded with the input image via channel concatenation, and CP2 dynamically interacts with the corresponding deep semantic features through CHKIM. 3) FEM: The fused features are convolved and upsampled to a uniform scale and concatenated in the channel dimension, and then AGFEM is used to adaptively enhance the features from different layers to produce the final predicted image.
• Figure 4: The architecture of MGMCB.
• Figure 5: Illustration of modulation modules. (a) Top-down global attentional modulation, (b) Bottom-up point-wise attentional modulation, (c) Framework diagram of the proposed CHKIM.