Research On Nondestructive Tire Defect Detection And Classification Using Deep Learning Technology

With the rapid development of the automotive industry and the continued growth of car ownership,the demand for tires has been greatly increased,and the quality of tires determines,to a large extent,the service life of tires as well as the protection for the lives and property of drivers during the driving process.The on-line quality inspection of tires also makes it easier for tire manufacturers to discover unqualified products in the production line timely,thereby to adjust the production process of mechanical equipment and even manufacturing technique of tires with the aim of avoiding the waste of production materials.Therefore,testing tire quality is an indispensable part of the tire production process,and has important research and economic significance for the development of the tire industry and road traffic safety.At present,many domestic and foreign scholars have proposed lots of non-destructive defect detection technologies.Among them,X-ray based tire defect detection technology has been rapidly developed in the tire industry and has achieved excellent results.However,due to the complex texture structure of tire X-ray images and the diversity of tire defects,tire non-destructive testing research is still very challenging for practitioners and researchers.This article uses tire X-ray images as detection objects,and adopts image analysis,pattern recognition,and deep learning techniques to develop a truly end-to-end tire defects automatic detection and classification algorithm.This paper firstly designs the tire defects detection and classification tasks with traditional image analysis and classification methods.Then,it replies on deep learning technology to design an end-to-end tire defect detection and classification algorithm.The end-to-end integrated algorithm framework is compared to the traditional detection and classification method via extensive experiment data analysis.The main innovations of this thesis are as follows:1)Aiming at the periodic complex texture features of tire X-ray images,a tire X-ray image defect detection method is proposed based on inverse transformation of principal component residual information.The main texture image without defect information is reconstructed by the first K principal components and their corresponding vectors.This reconstructed image is compared with the original image,then the defect image patches including their positions can be obtained after post image binarization and Morphological operations.The experimental results show that the algorithm can quickly detect defects of air bubbles and impurities in the tire.2)For the complex diversity of tire defects,it is difficult to model these defect features with a predefined fixed basis function.A tire image feature extraction algorithm with adaptive basis functions is proposed and applied to the automatic detection and identification for tire defects.Using ICA/TICA(Independent Component Analysis/Topological Independent Component Analysis)to learn the adaptive basis functions and filters from the tire defect samples,which these bases adapt to the characteristics of the defect image;then these adaptive filters are applied to extract the features of the defect images;finally,SVM is used to identify the defect types.The method is based on the unsupervised learning of defect sets and can extract the significant features of the defect images adaptively,besides the calculation is simple and can be processed in parallel.Experimental results show that this method has satisfactory recognition rate for classification for tire defects with complex shape and texture,and its classification accuracy rate is as high as 94.94%.3)In order to avoid the above traditional classification methods,the image features and classifier are designed artificially,and it is difficult to realize the perfect combination of the two.Advanced results have been achieved by means of deep learning on many benchmark data sets,also many difficult problems of artificial intelligence have been solved with it.Therefore this paper proposes to apply deep learning to the classification of tire defects.In the consideration of the contradictions between a limited number of labeled tire defect samples and deep learning requirements of large-scaled labeled training sets,the strategy of migration learning,data expansion,and network fine-tuning are used to avoid over-fitting.As a result,the classification recognition rate is as high as 96.51%.The experimental results verify the effectiveness of this method in tire defect classification tasks and provide an effective algorithm and reference for the automatic detection of tire defects in the course of actual production process.4)Due to high contrast variation among the tire defect images,the classification effect of the above-mentioned single-channel network will be affected.Inspired by the phenomenon of simultaneous release of microcortices in the cerebral cortex,a multi-channeled deep convolutional neural network is constructed.First of all,the algorithm designs a single-channel CNN network;then it constructs a multi-channeled convolutional neural network that consists of multiple single networks;also optimize the input data for each channel during network training with the aim to enhance its strong generalization ability and high robustness;subsequently determine the optimal number of channels for a multi-channeled convolutional neural network;finally analyze the time cost by comparing experimental results.Its recognition rate reaches as high as 98.47%.5)Aiming at the problem that the current algorithm is used to detect and classify tire defects in two separate stages,an algorithm based on the regional suggestion network is proposed to pack them as a whole.In view of the existence of multi-scaled changes in tire defects,the hierarchical neural networks with defect detection and classification are designed to position different scales for different output layers of neural networks.The experimental data shows that the detection rate is more than 97% except for the bubble defect detection rate of 91.56%,and this algorithm greatly improves the effect of small target defect detection.For the detection speed,the CNN feature upsampling is used instead of the image input upsampling method to reduce the storage and calculation costs and accelerate the detection speed.The experimental results show that the network can detect tire defects of all scales only by inputting a single-scaled image,and the detection speed can reach 10 fps when the image size is 1250*425 pixels on a workstation with 3.0 GHz Intel Core i7-5960 X CPUs and GeForce GTX 1080 8G Nvidia GPU,on an Ubuntu 14.04,Caffe and python 2.7 platform..