YOLO-L: A YOLO-Based Algorithm for Remote Sensing Image Target Detection

In response to the poor performance of object detection in remote sensing images under complex backgrounds, this paper proposes a new object detection algorithm model called YOLO-L based on the YOLOv5 algorithm. This algorithm model constructs a novel backbone network called BottleCSP, which refreshs the C3 module and SPP module in the original YOLOv5 backbone network. It integrates feature maps of different sizes to beef up contextual information, thereby better extracting deep linguistic information from images. The BottleCSP backbone network solves the problem of poor performance in handling images with small target recognition by customizing and adjusting the original Backbone backbone network. Additionally, the CoordAttention attention mechanism is inserted in both the Backbone backbone network and the Head prediction head, which helps the network to more accurately locate the objects of interest and enhance the model’s perception ability in different regions of the image. This paper conducts experiments on the RSOD public dataset, and compared with the original YOLOv5 algorithm and newer detection methods such as DC-SPP-YOLO, MRFF-YOLO, and YOLOv5-slight, the mAP (mean average precision) is improved by 12.3%, 16.22%, 4.47%, and 7.67%, respectively. This effectively improves the recognition performance of images with small targets.