DAC-Net: Divide-and-Conquer Network for Infrared Small Target Detection

Abstract

Infrared small target detection (IRSTD) plays a vital role in remote sensing and automated early-warning infrastructure. However, its performance is deeply constrained by the dim nature of targets and complex background clutter. Although convolutional neural networks (CNNs) have driven advancements in this field, they often struggle to capture global contextual information, leading to an inaccurate understanding of background semantics and triggering false alarms. Meanwhile, existing methods overlook the physical and morphological characteristics of the targets themselves, which can easily lead to boundary distortion and missed detections. To resolve these limitations, we propose a novel network based on a hybrid architecture, named DAC-Net. Specifically, we introduce the Edge-aware CNN-VMamba Fusion module into the encoder, which adopts a CNN-VMamba dual-branch structure to thoroughly learn both local details and global semantics. To enhance the network’s representation capability for the IRSTD task, a dynamic multiscale enhancement module is incorporated to effectively capture the physical and morphological characteristics of targets across diverse scales, achieving further feature selection and background suppression. Furthermore, an explicit edge loss is integrated into the joint optimization framework to supervise boundary gradient variations, thereby alleviating target shape distortion. Extensive experiments on NUAA-SIRST, NUDT-SIRST, and IRSTD-1K datasets demonstrate that our proposed DAC-Net outperforms existing methods across various evaluation metrics.