改进YOLOv7的玻璃表面缺陷检测方法研究  

Research on Improved Detection Method of Glass Surface Defect of YOLOv7

作　　者：宋吕明 刘明芹 李祥宾 朱雅 王家超 Song Lüming;Liu Mingqin;Li Xiangbin;Zhu Ya;Wang Jiachao(School of Ocean Engineering,Jiangsu Ocean University,Lianyungang,Jiangsu 222000,China;School of Mechanical Engineering,Jiangsu Ocean University,Lianyungang,Jiangsu 222000,China;Lianyungang Topu Technology Development Co.,Ltd.,Lianyungang,Jiangsu 222069,China)

机构地区：[1]江苏海洋大学海洋工程学院,江苏连云港222000 [2]江苏海洋大学机械工程学院,江苏连云港222000 [3]连云港市拓普科技发展有限公司,江苏连云港222069

出　　处：《机电工程技术》2024年第6期209-215,共7页Mechanical & Electrical Engineering Technology

基　　金：江苏省海洋资源开发研究院开放课题(JSIMR201810)。

摘　　要：针对电子产品玻璃边缘表面的电极区域与非电极区域深加工过程中产生的大划伤、小划伤、划痕、异物等缺陷,提出了一种基于改进的YOLOv7的玻璃表面缺陷小样本检测方法。首先,在主干网络中加入卷积注意力模块(Convolutional Block Attention Module, CBAM)提高了通道注意力与空间注意力,解决了玻璃表面缺陷面积较小、在图像中分布差异较大、提高卷积神经网络在缺陷区域学习鲁棒性的特征表示的问题。其次,考虑到工业生产过程中缺陷样本较少、样本量不均衡,采用随机高斯噪声、Mixup、随机填充图像和随机拼接等图像增强方法,将样本进行扩充,并使样本均衡化。最后,将增加一个预检测头用于细长且轻浅的划痕检测,结合其他3个预测头,四预测头结构可以有效缓解过大差异对象带来的尺度方差引起的负面影响。实验结果表明,改进的YOLOv7算法相较于原始算法,平均精度提高了6.15%(mAP),检测效果优于当前YOLOv7网络,在一定程度上提高了工业生产过程中玻璃表面缺陷的小样本检测精度。Aiming at the defects such as large scratches,small scratches,scratches and foreign bodies generated during the deep processing of electrode and non-electrode areas on the glass edge surface of electronic products,a small sample detection method for glass surface defects is proposed based on improved YOLOv7.First of all,convolutional block attention(CBAM)is added to the backbone network module to improve channel attention and spatial attention,solve the problem of small defect area on the glass surface,large distribution difference in the image,and improve the feature representation of the convolutional neural network learning robustness in the defect region.Secondly,considering the small number of defective samples and unbalanced sample size in the industrial production process,image enhancement methods such as random Gaussian noise,Mixup,random fill image and random splicing are adopted to expand and equalize the samples.Finally,a pre detection head is added for thin and light scratch detection.Combined with the other three prediction heads,the four-prediction head structure can effectively alleviate the negative impact of scale variance caused by excessively different objects.The experimental results show that compared with the original algorithm,the average accuracy of the improved YOLOv7 algorithm is increased by 6.15%(mAP),and the detection effect is better than that of the current YOLOv7 network,which can improve the detection accuracy of small samples of glass surface defects in the industrial production process to a certain extent.

关 键 词：YOLOv7 玻璃表面缺陷检测 卷积注意力模块 图像增强 四预测头结构 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]