MNHU-Net: A Multiscale Feature Fusion and Nested Structure-Based High-Order U-Net for Infrared Small Target Detection

Infrared small target detection (IRSTD) methods have been extensively investigated within the infrared search and tracking applications. U-shaped networks and their improved versions have significantly enhanced IRSTD segmentation performance in recent years. However, existing methods overlook the feature dilution and insufficient representation of long-range dependencies in feature maps, hindering the accurate segmentation of small targets obscured by background clutter. To handle this problem, we designed a multiscale feature fusion and nested structure-based high-order UNet (MNHU). Our approach utilizes a high-order UNet paradigm that progressively calibrates infrared feature maps, and integrates hierarchical features to enhance both small target and background texture extraction performance. High-order U-Net selectively integrates feature maps from highly correlated adjacent nodes. Subsequently, these encoding or decoding feature maps are progressively merged into a high-order fusion feature map, ensuring sufficient feature extraction and precise feature representation. We evaluate the proposed high-order UNet-based methods (MNHU-E, MNHU-D, and MNHU) on three public datasets. The results underscore our method's superior performance in target enhancement and texture awareness, outperforming state-of-the-art techniques in intersection over union, Dice similarity coefficient, precision, and sensitivity. MNHU exhibits robust segmentation capabilities and generalization, effectively achieving a balance between model complexity and performance, which highlights its suitability for practical applications.