RSS-YOLO: A Feature Fusion and Decoupled Object Detection Method in Remote Sensing Satellite Images

Itis still challenging to detect objects in optical remote sensing satellite imagery because of large-scale and small-scale objects coexisting in the images and variations in object density, making multi-scale object detection challenging. Additionally, there is a sample imbalance issue between easy and hard samples. This study proposes an improved approach based on the YOLOv7-Tiny framework to address these problems. The proposed RSS-YOLO algorithm introduces the concept of the Light Decoupled Head, which accelerates model convergence and enhances detection accuracy. Moreover, the Ghost-ASFF method is utilized to adaptively fuse cross-scale features with minimal parameter increase, thereby reducing the loss of object features. Further, a novel loss function enhances the gradient gain for hard samples, improving the accuracy and robustness of the network for object detection. The proposed RSS-YOLO algorithm achieves a mAP50 value of 69.3% on the DIOR dataset, surpassing typical detection algorithms such as SSD, Faster R-CNN and YOLO series. Additionally, it achieves a mAP50 value of 95.9% on the UCAS-AOD dataset, demonstrating superiority over SOTA object detection algorithms.