FDPF-Net: A Full-Scale Dynamic Pyramid Fusion Network for Infrared Small Target Detection

Infrared small target detection (IRSTD) methods have been extensively researched for various military and civilian applications and have greatly developed with the progress of deep learning in recent years. However, the performance of IRSTD remains limited due to challenges such as weak detection capabilities for diverse target boundaries and the complex background clutter present in infrared images across different scenarios. To overcome these challenges, this paper proposes a two-stage end-to-end full-scale dynamic pyramid fusion network (FDPF-Net). This network aims to refine small target boundary information and enhance both background consistency and the contrast between the target and its surroundings. The FDPF-Net introduces a feature extraction trunk sub-network and a full-scale dynamic refinement sub-network to extract and refine target and background information. Additionally, the proposed cross-layer scale adaptive module which is positioned between the trunk and the refinement sub-networks, adaptively integrates and optimizes the full-scale feature representation for boundary feature compensation. Finally, a feature pyramid fusion module is used to fuse and exploit the intrinsic information of small targets, avoiding feature dilution during the information passing process. Experimental results on three public datasets demonstrate that the proposed FDPF-Net outperforms other state-of-the-art methods in terms of Intersection over Union (IoU), Dice Similarity Coefficient (DSC), Precision (Pre), and Sensitivity (Se) and also exhibits more robust segmentation performance. It also maintains a balance between segmentation performance and model complexity, indicating its significant potential for real-world IRSTD applications.