Work

SCDE-Net: Strip-Context and Dilation-Enhanced Deep Unfolding Network for Infrared Small Target Detection

Year: 2026

Type: article

Abstract: Deep unfolding networks (DUNs) hold great potential for infrared small target detection (IRSTD) by embedding the domain prior of target-background separation into interpretable multistage architectures. However, the internal operators of representative unfolding architectures for IRSTD are still largely inherited from generic convolutional designs. Specifically, limited spatial perception makes it difficult to jointly handle structured directional clutter and local target contrast, while premature fusion of the separated background and target components weakens the discrimination available to subsequent stages. To address these issues, we propose a strip-context and dilation-enhanced deep unfolding network (SCDE-Net), which tailors the unfolding pipeline to the physical characteristics of infrared imagery. For background estimation, a strip context background module (SCBM) combines local convolution with orientation-aware strip pooling to suppress structured clutter while retaining local detail responses. For target recovery, a dilated context sparse module (DCSM) widens the receptive field via dilated convolutions to exploit local contrast cues. Furthermore, an adaptive residual fusion module (ARFM) retains the separated background and target representations as distinct inputs to learnable reconstruction, preserving component-specific cues for subsequent stages. Extensive experiments and visual analyses on three public benchmarks (IRSTD-1 K, NUDT-SIRST, and SIRST V1) demonstrate the effectiveness of SCDE-Net and illustrate the behavior of its unfolding stages.

Source: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

DOI: 10.1109/jstars.2026.3724296

Authors: Mengmeng Zhang, Dan DU, Hongyuan Jing, Huihui Bai

Institutions: Beijing Union University, Beijing Jiaotong University

Language: English

Topic: Infrared Target Detection Methodologies

Subfield: Aerospace Engineering

Field: Engineering

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $1,800

Funders: National Natural Science Foundation of China, Beijing Union University, Beijing Nova Program