基于YOLOv8优化改进的太阳能电池片缺陷检测模型  

A solar cell defect detection model optimized and improved based on YOLOv8

作　　者：彭自然 王思远 肖伸平[1,2] Peng Ziran;Wang Siyuan;Xiao Shenping(School of Electrical and Information Engineering,Hunan University of Technology,Zhuzhou,Hunan 412007,China;Hunan Key Laboratory of Electric Drive Control and Intelligent Equipment,Zhuzhou,Hunan 412007,China)

机构地区：[1]湖南工业大学电气与信息工程学院,湖南株洲412007 [2]湖南省电传动控制与智能装备重点实验室,湖南株洲412007

出　　处：《光电工程》2024年第11期88-101,共14页Opto-Electronic Engineering

基　　金：国家重点研发计划基金资助项目(2019YFE0122600);湖南省教育厅重点科研项目(22A0423);湖南省自科基金项目(2023JJ60267,2022JJ50073)。

摘　　要：针对太阳能电池片缺陷检测中存在检测精度低、误检和漏检率高的问题,本文在深度学习模型YOLOv8的基础上进行优化与改进,提出了一种太阳能电池片电致成像(electroluminescent,EL)缺陷检测模型。首先,采用自校准光照学习(self-calibrated illumination,SCI)方法对低光照图像进行预处理,以增强太阳能电池片缺陷的有效特征信息。然后,引入一个空间到深度的注意力模块(space-to-depth,SPD),替换主干网络的第二个跨步卷积层,避免跨步卷积导致的信息丢失,扩大感受野,减少计算量,从而在特征提取时保留更多特征信息。其次,构建了空间双向要素金字塔网络(spatial-BiFPN,S-BFPN),通过多尺度特征融合,解决因太阳能电池片缺陷形状多样性而造成缺陷识别率不稳定的问题。最后,本文改进了损失函数,使用MPDIoU作为损失函数,解决了原有的CIoU损失函数中惩罚项失效的问题。实验结果显示,改进后的YOLOv8模型的mAP达到了96.9%,比原始YOLOv8提升了2.2%,计算量减少了0.2 GFlops,检测速度最高达155 f/s,实现了高精度与高实时性,更适合工业部署。To address issues of low detection accuracy and high false-positive and false-negative rates in solar cell defect detection,this paper proposes an optimized solar cell electroluminescent(EL)defect detection model based on the YOLOv8 deep learning framework.First,a self-calibrated illumination(SCI)method is applied to preprocess low-light images,enhancing effective feature information for solar cell defects.Then,a space-to-depth(SPD)attention module is introduced,replacing the second stride convolution layer in the backbone network.This substitution avoids information loss caused by stride convolution,expands the receptive field,and reduces computational load,preserving more feature information during extraction.Next,a spatial-BiFPN(S-BFPN)network is constructed to perform multi-scale feature fusion,stabilizing defect recognition rates by addressing the shape variability of solar cell defects.Lastly,the loss function is improved by adopting MPDIoU,which resolves the issue of ineffective penalties in the original CIoU loss function.The experimental results show that the improved YOLOv8 model achieved an mAP of 96.9%,a 2.2%increase compared to the original YOLOv8.The computational load was reduced by 0.2 GFlops,and the detection speed reached a maximum of 155 f/s,demonstrating high accuracy and real-time performance,making it more suitable for industrial deployment.

关 键 词：深度学习 太阳能电池片 缺陷检测 YOLOv8 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]