SODRS: Semisupervised Learning for One-Stage Small Object Detection in Remote Sensing Images

Small object detection in remote sensing images faces challenges such as weak features, vulnerability to inter ference, and limited object visibility. These factors have made it a long-standing technical issue. Traditional object detection methods often perform poorly in this field, especially when dealing with high-resolution remote sensing images, where detec tion accuracy and robustness fail to meet practical application needs. Furthermore, due to the large volume of remote sensing image data and the high cost of annotation, effectively utilizing unlabeled data has become the key to enhancing model detection performance. To resolve these problems, this paper proposes an one-stage semi-supervised learning method for small object detection in remote sensing images (SODRS). This method uses FCOS as the baseline and introduces the Squeeze-and-Excitation (SE) attention mechanism to enhance feature representation, the Soft Focal Loss (SFL) module to optimize the class ambiguity between objects and background, and the Confident Pseudo Labeling (CPL) strategy to improve the quality of pseudo labels. Finally, the Conditional Random Fields (CRF)-based Label Refinement (CLR) is applied to post-process the predicted labels, improving spatial relationships and dependencies among objects. Experimental results demonstrate that SODRS excels at detecting small objects in complex remote sensing scenarios, with higher accuracy and robustness than existing classical methods. Notably, SODRS can accurately distinguish objects even when densely distributed, close to each other, or overlapping. This demonstrates the application potential of one-stage semi-supervised methods in small object detection in remote sensing images. The code for this method is available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/wxqnl/SODRS</uri>.