基于改进YOLOv8的街景图像变压器目标检测  

Transformer object detection in street view images based on improved YOLOv8

作　　者：廖方舟 杨晓霞[1] 杨容浩[2] 施琪琦 LIAO Fangzhou;YANG Xiaoxia;YANG Ronghao;SHI Qiqi(College of Geography and Planning,Chengdu University of Technology,Chengdu 610059;College of Earth and Planetary Science,Chengdu University of Technology,Chengdu 610059)

机构地区：[1]成都理工大学地理与规划学院,成都610059 [2]成都理工大学地球与行星科学学院,成都610059

出　　处：《电气技术》2024年第12期12-20,27,共10页Electrical Engineering

摘　　要：街景图像是一种城市街道级别信息地理大数据,利用街景图像不仅能够实现大范围、高效率的变压器巡检,还能降低巡检成本。但是,街景图像中的变压器往往像素少、分辨率低、背景复杂,导致现有目标检测方法对变压器的检测准确度不理想。针对上述问题,本文提出一种改进的YOLOv8算法YOLOv8-WSX。首先,使用WIoU作为损失函数,强化模型对困难样本的检测性能;然后,引入空间分组增强(SGE)注意力机制模块,提高模型的特征提取能力;最后,增加微小目标检测头,解决微小变压器目标漏检的问题。实验结果表明,相较于YOLOv8,YOLOv8-WSX的F1值提升了5.9个百分点,IoU阈值为50%时的平均精确率均值提升了6.3个百分点,IoU阈值在50%~95%范围内的平均精确率均值提升了3.2个百分点,且模型的参数量有所下降。Street view images are a form of geospatial big data at the urban street level.Utilizing street view images not only enables large-scale and efficient transformer inspection but also reduces inspection costs.However,transformers in street view images often have few pixels,low resolution and complex backgrounds,leading to unsatisfactory precision of existing object detection methods.To address these issues,this paper proposes an improved YOLOv8 algorithm named YOLOv8-WSX.Firstly,wise intersection over union(WIoU)is used as the loss function to strengthen the detection performance of the algorithm for difficult samples.Secondly,the spatial group-wise enhance(SGE)attention mechanism module is introduced to improve the feature extraction ability of the algorithm.Finally,an extra-small object detection head is added to solve the problem of missing detection of extra-small transformer objects.The experimental results show that compared to YOLOv8,YOLOv8-WSX increases the F1 score by 5.9 percentage points,increases the mean average precision by 6.3 percentage points for IoU at 50%,and increases the mean average precision by 3.2 percentage points for IoU from 50%to 95%.Additionally,the model has fewer parameters.

关 键 词：变压器目标检测 YOLOv8 深度学习 街景图像 

分 类 号：TU984.113[建筑科学—城市规划与设计] TM41[电气工程—电器] TP391.41[自动化与计算机技术—计算机应用技术]