基于卷积神经网络的轮胎缺陷X光图像分类  

Defect classification for tire X-ray images using convolutional neural network

作　　者：崔雪红[1] 刘云[1] 王传旭[1] 李辉[1] 

机构地区：[1]青岛科技大学信息科学技术学院,青岛266061

出　　处：《电子测量技术》2017年第5期168-173,共6页Electronic Measurement Technology

摘　　要：轮胎缺陷的类型直接决定着轮胎是否为残次品或废品,对于轮胎定级具有重要参考价值,探索高性能的轮胎缺陷分类方法至关重要。采用卷积神经网络技术,提出一个端到端的图像自动分类算法。首先,从国内某轮胎生产线上通过现场运行的轮胎X光射线缺陷检测系统采集胎侧异物缺陷、胎冠异物缺陷、气泡缺陷、胎冠劈缝、胎侧开根5种最常见缺陷类型和1种正常胎侧图像作为分类目标,并且依据缺陷图像的缺陷尺度,将每幅图像缩放到127×127像素的统一大小;然后,设计含有5个卷积层、3个池化层、3个全连接层的网络结构。最后,用采集的缺陷样本对所设计的深度网络进行训练学习与测试。并将该算法和大量传统分类算法进行实验对比,取得更好的分类效果,平均测试识别率高达96.51%。Type of tire defects directly determines whether the tire is defective products or waste, which has important reference value for tire grading, it is vital to explore high performance tire defect classification method. First, collecting five common types of defects and a normal images from a typical tire manufacturing, namely belt-foreign-matter, sidewall-foreign-matter, belt-joint-open, cords-distance, bulk-sidewall and normal-cords, was used to perform the tire defect classification experiments. And down-sampling or up-sampling the images in the dataset to a fixed resolution of 127×127. And then designing depth network which contains 5 convolutional layers, 3 max-pooling layers, 3 fully- connected layers. Finally, training and testing the designed depth network with defect samples collected. Experimental results showed that the method proposed has higher recognition rates for tire defects than other algorithms, the averaged rate of recognition is high to 96.51%.

关 键 词：卷积神经网络 池化 轮胎缺陷 图象分类 全连接 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN081[自动化与计算机技术—计算机科学与技术]