电致发光缺陷图像非均衡样本生成与评估  

Generation and Evaluation of Imbalanced Samples in Electroluminescent Defect Images

作　　者：沈熠辉 颜胜男 刘华锐 游宏亮 SHEN Yihui;YAN Shengnan;LIU Huarui;YOU Hongliang(School of Arts and Sciences,Fujian Medicine University,Fuzhou,China,350122;Department of Mathematics and Data Science,Minjiang University,Fuzhou,China,350108;Department of Computer Science and Technology,Minjiang University,Fuzhou,China,350108;National Photovoltaic Industry Metrology Center,Fujian Academy of Metrology,Fuzhou,China,350003)

机构地区：[1]福建医科大学文理艺术学院,福州350122 [2]闽江学院数学与数据科学学院,福州350108 [3]闽江学院计算机科学与技术系,福州350108 [4]福建计量科学研究院国家光伏产业计量测试中心,福州350003

出　　处：《福建电脑》2023年第10期16-20,共5页Journal of Fujian Computer

摘　　要：太阳能电池板缺陷降低了光电转换效率和使用寿命,因此必须进行缺陷检测。YOLO和GoogleNet等深度学习模型提高了准确性,但数据集的缺陷样本不平衡会影响模型的准确度。为解决此问题,本文提出了改进的扩散概率模型。该模型使用sigmoid方案更新加噪策略,生成缺陷数据中的黑斑片,并通过目视判断和目标检测算法评价生成图像质量。实验结果显示,与传统数据增广、DCGAN和DDPM等方法相比,SIG-DDPM在生成图像质量方面更出色。经过目视判断和目标检测算法的测试,生成的黑斑电池片图像接近真实图像,说明了该方法的有效性。Defects in solar panels reduce the photovoltaic conversion efficiency and lifespan,making defect detection essential.Deep learning models like YOLO and GoogleNet have improved the accuracy of defect detection.However,the imbalanced distribution of defect samples in the dataset affects the model's accuracy.To address this issue,this article proposes an improved diffusion probability model.This model uses the sigmoid scheme to update the denoising strategy,generate black patches in defect data,and evaluate the quality of the generated image through visual judgment and object detection algorithms.The experimental results show that SIG-DDPM performs better in generating image quality compared to traditional data augmentation,DCGAN,and DDPM methods.After visual judgment and target detection algorithm testing,the generated black spot battery image is close to the real image,indicating the effectiveness of this method.

关 键 词：扩散概率模型 电致发光缺陷 图像生成 图像评价 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]