摘要
基于高密度遮挡下鸟巢识别正确率低的问题,提出一种基于Transformer技术改进的Yolov5变电站鸟巢识别算法模型。首先将现有Yolov5算法原本的Yolo预测头(Yolo Prediction Heads)替换为Transformer预测头(Transformer Prediction Heads,TPH),同时,为了提高对小物体的检测能力新增了一个预测头。然后嵌入卷积注意力模块(Convolutional Block Attention Module,CBAM),提取注意区域,以抵制干扰信息,关注有用的目标对象。最后利用CSPDarknet53结构设计思想,提取最终的特征图,经CSP模块将输出的特征图进行快速降维。实验结果表明,对比Yolov5算法,TPH-Yolov5算法的mAP(Meanaverage Precision)值提高了15.7%。
Aiming at addressing low accuracy problem of bird nest recognition under high density occlusion,a modified Yolov5 algorithm model is proposed.Firstly,the original Yolo prediction heads of the existing Yolov5 algorithm is replaced with transformer prediction heads(TPH),and a new prediction head is added to improve the detection of small objects.Then,CBAM is embedded to extract atte ntion areas,so as to resist confusing information and focus on useful target object.Finally,the final feature map is extracted using the structure design idea of CSPDarknet53,and the dimension of output feature map is rapidly reduced through CSP module.Experimental results show that mAP value of the TPH-Yolov5 algorithm is increased by 15.7%compared with Yolov5 algorithm.
作者
缪苗
袁峰
邹明翰
任圣雄
MIAO Miao;YUAN Feng;ZOU Minghan;REN Shengxiong(State Grid Nanjing Power Supply Company,Nanjing 210019,China)
出处
《电工技术》
2023年第21期58-62,共5页
Electric Engineering
