MultiScale-enhanced detection network (MS-EDN) with dual encoder structure for infrared small target detection

Abstract

Infrared small target detection is vital in military, security, and rescue operations. While deep learning has achieved remarkable progress in general detection frameworks, its application to infrared imagery remains constrained by intrinsic feature representation challenges. Existing methods are often affected by complex backgrounds, causing small targets to be overlooked, especially in low-contrast environments where detailed information is easily lost. This paper proposes a dual encoder network with multi-scale enhanced detection to address these challenges. One branch incorporates a feature residual enhancement module that combines a residual convolutional block attention module with a feature enhancement module for efficient feature extraction. The other branch integrates a dynamically parallelized patch-aware attention module and employs a multi-branch extraction strategy to capture information across various scales. The multi-scale dynamic fusion module in the neck layer enhances feature representation, facilitating accurate detection and localization of small targets. Additionally, soft pooling is used in the downsampling process to better preserve important features while reducing information loss. Extensive experiments on the SIRST dataset demonstrate that the proposed method outperforms existing approaches in effectiveness and robustness, effectively extracting small target information in complex backgrounds.