基于PPYOLO的电池片缺陷检测  

Defect detection of solar cells based on PPYOLO

作　　者：韩钰[1] 郑金亮[1] 王磊[1] 蔡培君[1] 王晨旸 王紫玉 HAN Yu;ZHENG Jinliang;WANG Lei;CAI Peijun;WANG Chenyang;WANG Ziyu(Jianghuai College of Anhui University,Hefei 230031,China)

机构地区：[1]安徽大学江淮学院,安徽合肥230031

出　　处：《邵阳学院学报（自然科学版）》2023年第4期8-16,共9页Journal of Shaoyang University：Natural Science Edition

基　　金：安徽省高等学校科学研究重点项目(2022AH053061)。

摘　　要：为进一步提升太阳能电池片表面缺陷检测效率,改进深度学习检测算法PPYOLO,添加索贝尔(Sobel)算子对输入图片进行处理得到缺陷边缘信息,此边缘信息会过滤大部分正常区域信息,同时能更快找到缺陷区域位置。再将边缘信息和原始图片拼接输入到网络模型中,这样既保留了原始图片的特征,又能防止Sobel算子处理后的信息存在缺失所导致的特征信息丢失。为验证该算法,在高清电池片数据集上进行仿真实验。实验结果表明,添加Sobel算子后的改进PPYOLO模型检测效率有所提升,表现为mAP提升1.4%,并且mAP相比于其他不同输入尺寸和模型大小的算法提升效果超过20%。In order to improve the detection efficiency of solar cell surface defects,the deep learning detection algorithm PPYOLO was improved,and Sobel operator was added to process the input image to obtain the defect edge information.The edge information will filter most of the normal area information and can find the defect area location faster.Then the edge information and the original image were combined to input into the model,which not only preserved the features of the original image,but also prevented the loss of feature information caused by the missing information processed by Sobel operator.In order to verify the algorithm,a simulation experiment was carried out on high-definition battery data.The experimental results show that the improved PPYOLO model detection efficiency is improved after adding Sobel operator,and the mAP increases by 1.4%.Compared with other algorithms with different input sizes and model sizes,the improvement effect of mAP is more than 20%.

关 键 词：缺陷检测 电池片 PPYOLO 索贝尔(Sobel)算子 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]