摘要
针对电致发光条件下太阳电池缺陷检测过程中缺陷尺度不一、背景噪声干扰严重的问题,提出一种基于YOLOv8的缺陷检测算法,以精确识别电池表面的缺陷。首先,采用感受野注意力卷积使模型动态调整感受野权重,减少图像信息的模糊和遗漏,提升对微小缺陷的检测效果;其次,引入坐标注意力机制,获取长宽方向信息,突出重点特征,减轻背景噪声干扰;最后,使用SIoU作为边界框损失函数,在预测框收敛过程中将向量角度纳入考量标准。实验结果表明:该算法在太阳电池缺陷检测数据集上mAP达到92.83%,相比于原始网络提高2.1%,误检、漏检的情况明显减少,能有效定位和识别太阳电池上的缺陷。
Aiming at the problems of defect scale inconsistency and serious interference of background noise in the defect detection process of solar cells under electroluminescence conditions,a defect detection algorithm based on YOLOv8 is proposed to accurately identify the defects on the surface of the cells.Firstly,the sensing field attention convolution is used to make the model dynamically adjust the sensing field weights,reduce the blurring and omission of image information,and improve the detection effect of tiny defects.Secondly,the coordinate attention mechanism is introduced to obtain the aspect direction information,highlight the key features,and alleviate the interference of the background noise;finally,the SIoU is used as the bounding box loss function,and the vector angle is incorporated into the consideration criterion in the process of predicting the convergence of the box.The experimental results show that the algorithm achieves a mAP of 92.83%on the solar cell defect detection dataset,which is an improvement of 2.1%compared with the original network,and the cases of misdetection and omission are significantly reduced,which is able to effectively locate and identify the defects on solar cells.
作者
王宗良
陆丽
康小东
阮晓鹏
笪贤文
Wang Zongliang;Lu Li;Kang Xiaodong;Ruan Xiaopeng;Da Xianwen(School of Electrical Engineering,Shanghai Dianji University,Shanghai 201306,China)
出处
《太阳能学报》
北大核心
2025年第8期379-386,共8页
Acta Energiae Solaris Sinica
基金
国家自然科学基金(42176194)。
关键词
深度学习
太阳电池
电致发光
卷积神经网络
目标检测
deep learning
solar cells
electroluminescence
convolutional neural networks
object detection
作者简介
通信作者:陆丽(1974-),女,博士、副教授,主要从事模式识别与智能系统技术方面的研究。lul@sdju.edu.cn。
