YOLO-GCRS: A Remote Sensing Image Object Detection Algorithm Incorporating a Global Contextual Attention Mechanism

Abstract

With the significant advancements in deep learning technology, the domain of remote sensing image processing has witnessed a surge in attention, particularly in the field of object detection. The detection of targets in remotely sensed images is a challenging task, primarily due to the abundance of small-sized targets and their multi-scale distribution. These challenges often result in inaccurate object detection, leading to both missed detections and false positives. To overcome these issues, this paper presents a novel algorithm called YOLO-GCRS. This algorithm builds upon the original YOLOv5s algorithm by enhancing the feature capture capability of the backbone network. This enhancement is achieved by integrating a new module, the Global Context Block (GC-C3), with the C3 backbone network. Additionally, the algorithm incorporates a convoluted block known as CBM (Convolution + BatchNormalization + Mish) to enhance the network model’s capability of extracting depth features. Moreover, a detection head, ECAHead, is proposed, which integrates an efficient attention channel (ECA) for extracting high-dimensional features from images. It achieves higher precision, recall, and mAP@0.5 values (98.3%, 94.7%, and 97.7%, respectively) on the publicly available RSOD dataset compared to the original YOLOv5s algorithm (improving by 5.3%, 0.8%, and 2.7%, respectively). Furthermore, when compared to mainstream detection algorithms like YOLOv7-tiny and YOLOv8s, the proposed algorithm exhibits improvements of 2.0% and 7.5%, respectively, in mAP@0.5. These results provide validation for the effectiveness of our YOLO-GCRS algorithm in addressing the challenges of missed and false detections in remote sensing object detection.