基于改进YOLOv5的太阳能电池板缺陷检测算法
Defect detection algorithm for solar panels based on improved YOLOv5
朱栋 1贺森1
作者信息
- 1. 常州大学微电子与控制工程学院 常州 213159
- 折叠
摘要
为提高太阳能电池板缺陷的检测精确,提出了一种改进的YOLOv5网络,对太阳能电池板常见的划痕、叉隐、黑斑、黑边以及无电等5类主要缺陷进行检测和分类.首先,使用改进后的ODConv模块对主干提取网络中的普通卷积模块进行替换,减少网络模型的参数量;其次,将C3模块中的Bottleneck结构替换成包含ParNet模块的Res2Net以增加感受野,从而提升了探测物体缺陷的能力和检测精确;最后,在预测网络前引入自适应特征融合结构,以融合不同特征图的位置与类别信息,增强特征表达并提高模型的鲁棒性.对自建的数据集进行训练、验证以及测试,实验结果表明,改进后的模型能够成功识别和定位5类常见缺陷.与原YOLOv5算法相比,在保持原网络高效性的同时,平均检测精确提升了6.2%.
Abstract
To enhance the accuracy of detecting defects in solar panels,an improved YOLOv5 network is proposed for the detection and classification of five common types of defects:scratches,cross hatching,dark spots,black edges,and no electricity.Firstly,the conventional convolutional modules in the backbone network are replaced with an improved ODConv module to reduce the model's parameter count.Secondly,the Bottleneck structure in the C3 module is replaced with a Res2Net containing the ParNet module to increase the receptive field,thereby improving the capability to detect object defects and the overall detection accuracy.Lastly,an adaptive feature fusion structure is introduced before the prediction network to fuse the position and category information of different feature maps,enhancing feature representation and improving the model's robustness.The model is trained,validated,and tested on a custom dataset,and experimental results demonstrate that the improved model successfully recognizes and locates the five common defects.Compared to the original YOLOv5 algorithm,the average detection accuracy is increased by 6.2%,while maintaining the efficiency of the original network.
关键词
缺陷检测/ODConv/ParNet模块/Res2Net/特征融合Key words
defect detection/ODConv/ParNet Block/Res2Net/feature fusion引用本文复制引用
基金项目
江苏省科技支撑项目(DFJH202131)
出版年
2024
NETL
NSTL
维普
万方数据