YOLO-SDH: improved YOLOv5 using scaled decoupled head for object detection

As a crucial component of computer vision tasks, object detection serves a significant role in a variety of fields, including autonomous driving, defect detection, and remote sensing image recognition. However, the majority of current object detection networks fail to achieve a decent balance between detection accuracy and detection efficiency, and there is room for improvement in terms of detection accuracy. In response, to improve detection accuracy, a more efficient network framework, YOLO-SDH, was proposed in this paper based on You Only Look Once v5 (YOLOv5). In addition, a decoupled head that automatically adjusts the number of channels according to the model size was proposed, which can enhance the network’s detection effect by separating the classification and regression tasks.On the premise of requiring less computation, a lightweight deformable convolution module is proposed so that the convolution can extract ROI over a wider range, thereby enhancing the accuracy of the object detection network. Experiments were run on the datasets of PASCAL VOC2012, NEU-DET, AW, and RSOD. In comparison to the original YOLOv5, the mAP 0.5 of YOLO-SDH improved by 1.29–3.03%, the F1-score improved by 1.2–3.2%, the Precision improved by 0.7–4.2%, demonstrating the algorithm’s efficacy and superiority.