An efficient training method for SAR object detection based on adaptive sample quality

In time-sensitive applications such as satellite reconnaissance and disaster response, fast and accurate detection of targets in Synthetic Aperture Radar images is of critical importance. This places dual demands on object detection models for both efficient training and rapid deployment. However, the presence of pervasive speckle noise, unstable image quality, and the complex relationship between targets and backgrounds in SAR images often leads to slow convergence and limited performance when using conventional training methods. To address these challenges, this paper proposes an efficient training method for SAR object detection based on adaptive sample quality, aiming to enhance the learning efficiency and final performance of detection models within limited training time. Experimental results show that, on the HRSID dataset, the proposed method improves mAP@50 by 0.162, 0.117, and 0.109 compared to the Random, CLF-RSD, and CLML strategies. On the SAR-Aircraft dataset, the improvements are 0.189, 0.161, and 0.124.