LD-Net: A Progressive Fusion Network for Enhanced Target Detection in Synthetic Aperture Radar Images

Synthetic Aperture Radar (SAR) technology plays a crucial r ole in target detection across military and civilian domains d ue to its all-weather capabilities and strong penetration. To a ddress the challenges posed by significant target scale variati ons in SAR image target detection, we propose a novel light weight deep learning framework that integrates multiple mo dules for enhanced detection accuracy and efficiency. The m odel utilizes (1) the LD-net, a feature fusion network designe d to mitigate information loss by effectively combining low-and high-level feature representations. (2) the The content-a ware reassembly operator (CARAFE) operator for content-a ware upsampling, which improves feature retention and utili zation; (3) the Receptive Field Block (RFB) module, simulating the human visual system's receptive field to better captur e detailed target features; and Experimental validation on th e SAR-AIRcraft-1.0 dataset demonstrates that our model sig nificantly outperforms baseline approaches, with an approxi mate 1.6% improvement in accuracy over the benchmark Y OLOv8 model, confirming its effectiveness for multi -scale S AR target detection.