基于CNN的工件缺陷检测方法研究噩-计算机科学.PDFVIP

第 44 卷第11A 期 Vo l. 44 No. 11A 计算机科学 2017 年 11 月 COMPUTER SCIENCE Nov.2017 基于 CNN 的工件缺陷检测方法研究 乔丽赵尔敦刘俊杰程彬 (华中师范大学计算机学院 武汉 430079) 摘 要 将卷积神经网络(CNN)应用于工件缺陷检测，来检测工件在生产过程中表面出现的缺陷，以提高工件的生 产质量。利用 CNN 可以对工件的图案进行识别，但识别无法检测出细微缺陷的问题。在 CNN 进行工件图案识别的 基础上，研究一种利用 CNN 实现缺陷检测的方法。该方法通过扩充缺陷样本，利用 CNN 识别的中间输出参数，定义 了缺陷分辨率的概念来衡量缺陷的程度，当缺陷分辨率达到一定水平时则认为是元缺陷图案，否则认为其存在缺陷。 实验验证了提出的 CNN 工件缺陷检测方法的有效性，数据表明缺陷检出率可达到 93.3% 。 关键词 工件缺陷检测，卷积神经网络，训练样本，缺陷分辨率 中图法分类号 TP39 1. 43 文献标识码 A Research of Workpiece Defect Detectioo Method ßased 00 CNN QIAO Li ZHAO Er-dun LIU Jun-jie CHENG Bin (Co!lege of Computer ,Central China Normal University ，引Tuhan 430079 ,China) Abstract The application of convolutional neural network (CNN) was proposed in workpiece defect detection ,which can detect workpiece defects on its surface ,to improve the product quality. The CNN can t recognize the small defects of a product although it can classify different objects very wel l. This paper presented a method which uses CNN for de fect detection based on the results of a recognition process. Firstly the defective samples are expanded to overcome the difficulty for lacking of training samples. Then by observing the output data obtained from a recognition CNN ,a concept called the defect distinguish ratio is defined to measure the degree of defection. It is considered as a non-defect pattern only when the defect distinguish ratio reaches a certain level. Finally ,the experiment demonstrates the validity and feasi- bility of the method ,in which the defect detection ratio can reach 93. 3%. Ke归lords Workpiece defect detection ,CNN ,Traini