A Novel Object Detection Algorithm for UAV Aerial Images Based on YOLOX_s

In this paper, a YOLOX_s based algorithm is used for object detection in UAV aerial images. It addresses challenges such as small objects, fuzzy features, and complex backgrounds. First, YOLOX _s algorithm inherits the advantage of feature pyramid structure of CSP-Darknet53 and utilizes the fusion of different scale feature maps to improve the detection performance of the network for different scale targets; Second, it separating anchor frame prediction from feature extraction with Decoupled Head, better utilizing the semantic information between the underlying and higher level features; finally, Anchor-Free and Advanced label assigning strategy SimOTA enable the network to obtain better training efficiency and detection accuracy. Simultaneously, it is able to plan the attribution of the targets from the global perspective, which to a certain extent solves the problem of the fuzzy and overlapping anchor frames. Experimental results show that the improved algorithm achieves 37.5 mAP on COCO dataset, outperforming popular algorithms like SSD, yolov3, yolov8, etc., by an average of 23.46% mAP.