基于改进YOLOv8s的多尺度检测算法OA北大核心CSTPCD

In view of the low detection accuracy caused by insufficient feature information of small objects like insulators and the large model size that is not conducive to hardware transplantation,a multi-scale detection algorithm MPH-YOLO(multiple prediction head-you only look once)is proposed.The algorithm MPH-YOLO improves the detection ability of small objects by expanding the detection scale of small object first,and then the CIoU loss function in YOLOv8s is replaced with the SIoU loss function as the edge loss function,so as to enhance the positional accuracy of the object.The Ghost convolution with lower cost is introduced instead of the traditional convolution in the network structure,so as to lighten the volume of the model.In terms of the improved algorithm,both its detection accuracy and model lightweight on the insulator dataset are improved,with a detection accuracy of 86.2%for mAP50-95 and a model volume of only 4.7 MB.The experimental results show that the MPH-YOLO not only can improve the detection of small objects effectively,but also be more lightweight for hardware transplantation,so it has high practical value.