Research On Tire Defect Detection Using Deep Learning Techniques

With the development of the economy and the improvement of people’s living standards,cars have become an indispensable transportation for people to work and travel.Tire,as a part of the car that directly contacts the ground,is mainly used to carry the weight of the car,cushion and absorb shock.Tire quality is extremely important to ensure the safety of drivers in safe driving.Traditional tire quality inspection mainly relies on X-ray imaging and artificial naked eye inspection,which has some disadvantages including strong subjectivity,low inspection efficiency,and poor real-time.In recent years,with the rapid development of computer vision technology,domestic and foreign scholars have proposed a series of industrial defect detection algorithms,which provide a reference for the surface quality evaluation of industrial products.However,due to the complex anisotropic texture background of tire X-ray images,the low visual quality between defects and the background,the similarity between various types of defects,and the differences between similar defects,which pose a huge challenge for traditional tire defect detection method.This paper takes X-ray images of common passenger car radial tires(hereinafter collectively referred to as tires)as the research object,and proposes methods for classification,detection and segmentation of tire defects based on deep learning technology.Comparative experiment with the traditional method is designed to verify the superiority of the method in this paper.The main research contents and contributions of this paper are as follows:(1)The tire defect dataset was established,covering the main defect types of tire X-ray images,which is composed of 1450 images in six different defect categories,including 337 sidewall-foreignmatter defects,245 tread-foreign-matter defects,218 tread-cord-cracking defects,132 sidewall-cordoverlap defects,331 sidewall-cord-cracking defects,and 187 bubble defects.LabelImg and Labelme applets were used to make corresponding target detection rectangular box labels and semantic segmentation pixel-level labels respectively.Random rotation,scale transformation,pixel translation,and stretching were utilized to expand the dataset,and the expanded samples reached 7,250.(2)A tire defect classification algorithm based on CNN is proposed.Traditional object classification methods based on feature extraction and classifiers mostly require empirical and subjective setting of parameters,which leads to the divergence of classification results.To overcome the ill-posedness of traditional classification methods for defect classification under multi-texture and anisotropic backgrounds,this paper proposes a method based on convolutional neural networks,which extracts the depth features of defects and classifies tire defects,transfer learning strategies and optimizers were used to accelerate the convergence of the network.The experimental results show that the comprehensive classification accuracy of the classification method proposed test set reach96.2% on test dataset,and the average classification time of the test set image is 0.012s/image.(3)In view of the mainstream object detection deep neural network,the object loss phenomenon is prone to occur in the detection process.In order to solve the problem of small object defect detection under the background of complex tire texture,increase the adaptability of object detection network for small object defects.Based on the YOLOv3 network,this paper proposes a network anchor box size optimization method based on the K-means clustering algorithm to increase the network’s ability for tire defects.For the object detection network IoU(Intersection over Union,IoU)loss function,when IoU = 0,the distance between the predicted bounding box and the real box cannot be reflected,and there is no gradient return,and training cannot be performed.The C detection box is introduced,which is the smallest rectangular box containing the predicted bounding box and the real box,which solves the problem of training abort when the predicted bounding box and the real box do not overlap.The above method was used to optimize the YOLOv3 network.The experimental results show that the AP(Average Precision,AP)value of the optimized YOLOv3 reaches 91.39%,achieving high-quality region of interest extraction of tire defects.the resolution is 2469*11400.The inspection time for the entire tire X-ray image with a resolution of 2469*11400 is 1.090 s,which can meet the online inspection requirements of industrial production.(4)To solve the problem that the low-level and high-level features of the feature extraction network have strong characterization capabilities for the shape,edge and semantic information of the defect respectively,the high-level features alone cannot fully reflect the contour information of the defect.Based on PSPNet(Pyramid Scene Parsing Network,PSPNet),pyramid feature fusion network was utilized to achieve the fusion of high-level and low-level features,combined with spatial pyramid pooling to alternately expand the receptive field.At the same time,an attention mechanism was introducted into feature extraction network for better characterization useful features,weaken useless features.The improved network model has achieved good results in tire defect segmentation.The segmentation accuracy mIoU(Mean Intersection over Union,mIoU)value is 87.86%,the average detection time of the test set image is 0.068s/image,and the detection time of the entire tire image is1.158 s,which can be achieved efficient product intelligent quality control.