Low resolution remote sensing object detection with fine grained enhancement and swin transformer

Abstract

Abstract Object detection in remote sensing images is a highly complex and challenging task. Remote sensing images typically suffer from issues such as small target sizes and densely distributed targets. Existing object detection algorithms often underperform in such scenarios due to their limited capability in handling fine-grained details and multi-scale objects. To address the persistent challenges in remote sensing image object detection, this study introduces a novel detection framework comprising three key innovations. First, we propose the Fine-grained Enhanced Downsampling Network (FEDNet) as the feature extraction backbone, specifically designed to preserve critical target information during downsampling through enhanced fine-grained feature representation. Second, we develop the Swin Transformer-based Progressive Aggregation Network (STPANet), which integrates Swin Transformer Blocks into the C3CST module to achieve superior multi-scale feature fusion while simultaneously capturing global contextual information and local spatial details. Finally, we incorporate the Shape-IoU loss function to optimize bounding box regression, significantly improving small target detection accuracy while maintaining computational efficiency. Experimental results demonstrate that the proposed method achieves outstanding performance on the DOTA and DIOR datasets, with mean average precision (mAP@50) scores of 69.9% and 85.5%, respectively. These results highlight its superior detection performance under low-resolution conditions.