| Tires are one of the key components of automobiles,and their quality is of great significance to driving safety.The industrial production precision of radial tire is strict and the steps are complicated.Different types of defects will appear,which will affect the use of tires and cause hidden dangers to the public lives.Therefore,to pledge the tire security during driving,it needs to execute strict quality inspection before leaving the factory.This work studies some defects of TBR tire.The research content mainly includes four parts:(1)The preprocessing of tire X-ray image data before entering network training.Firstly,the image is splited to windows to reduce the amount of calculation;then,the median filter algorithm is applied to operate the noise.At the same time,the image data is enhanced by histogram equalization.Secondly,to alleviate the over fitting problem of the training network,the tire X-ray image data set is expanded through geometric transformations to increase the size of the data set.Finally,the Label Me toolkit is used to label the preprocessed images and generate a data set for supervised learning.(2)Faster R-CNN network model is applied to detect the defects of X-ray images of industrial tires.The experimental results show that although the algorithm can locate the tire defects,it can only locate part of the area,which has the phenomenon of false detection and missing detection.Especially,the weak edge bubble defects in the X-ray image have poor positioning and recognition effect,which can not attain the needs of industrial detection.Therefore,it needs improving the algorithm.(3)Three improvements are made to the tire defect detection algorithm based on Faster R-CNN.Firstly,the ZF network structure only uses the fifth layer of convolutional features for positioning.The third layer of convolution output and the fifth layer of convolution output are combined as the input of the regional suggestion network layer.Secondly,After the Ro I pooling layer,the O nline Hard Example Miniing(OHEM)algorithm is introduced,and the selected high-loss samples are used for training in stochastic gradient descent.Finally,according to the defect characteristics in the tire X-ray image and the size of the hand-marked frame,re-select the anchor scale and the aspect ratio.(4)The superiority of the algorithm are analyzed in three aspects.Firstly,it contrasts to the Faster R-CNN algorithm and the mend algorithm,the accuracy of this algorithm is higher,and it is found that the value becomes larger,indicating that the detection accuracy of the algorithm in this paper is improved.Then,the accuracy,test time and training time before and after the algorithm improvement are compared and analyzed to verify that the algorithm can attain the real-time detection requirements of Industry 4.0.Finally,contrast to other advanced deep learning networks,the performance of the algorithm to further verify the detection effect. |
