Research On Tire Non-destructive Testing Method Based On Semantic Segmentation

Tire quality affects vehicle and driver safety.The inner defect of tire is the key for the driving safety.X-ray imaging is used for non-destructive testing of tires.Previous methods rely mainly on manual inspection,which are inaccurate and subjective.With the development of computer vision technology,its application in industrial production is becoming more and more popular,which greatly improves the accuracy and efficiency of quality assessment.However,tire radiographic image has complex texture background,the contrast between the defect and the texture background is low,and the tire defect has the characteristics of inter-class similarity and intra-class diversity.This makes the traditional visiual inspection technology difficult to perform the task of tire defect detection.How to quickly and accurately locate defects and describe their geometric features is a problem to be solved in the field of tire defect detection,which is of great significance for tire manufacturers to improve product quality.This dissertation researches on tire defect detection in radiographic image,and proposes two methods based on semantic segmentation using deep learning technology.The effectiveness of the proposed methods are proved by comparing with the current state-of-the-art algorithms.The main contributions of this dissertation are as follows:(1)A multi-scale aggregation network based on attention guidance is proposed for pixel-wise tire defect detection.Because of the complex texture background of the tire radiographic image,it is challenging for the common convolution neural network to distinguish the difference between the defective region and the texture background,resulting in poor geometric feature representation.To solve this problem,a multi-scale aggregation network based on attention guidance is proposed.First,Res Net50 is used as the backbone feature extraction network to extract feature information.The proposed spatial aggregation pyramid module extracts multi-scale information of high-level features,and uses attention refinement branches to select effective defect features.Second,the proposed dual attention upsampling module can gradually restore the edge details of defect features,while taking into account the semantic information of features.Finally,to obtain the feature map dependencies between channels,a semantic enhancement module is proposed to aggregate multi-scale features.The channel feature matrix provides the feature representation of specific semantics,guides the enhancement of features,and better represents the defect features.Ablation experimental results show that the proposed modules can effectively improve the performance of the proposed network.The comparative experimental results with thestate-of-the-art segmentation networks further validate that the proposed network can achieve satisfactory performance in terms of segmentation accuracy and classification accuracy,achieving89.13% and 98.65%,respectively.Moveover,the average detection time of achives0.0280 s,which can can implied on practical intelligent quality control of products.(2)Considering the inherent inductive bias in convolutional neural network,which cannot model the long distance pixels in the tire X-ray image,and lacks the ability to capture the global information of the image.A lightweight Transformer-based defect detection network is proposed for tire inspection,in which the self-attention mechanism in Transformer can capture long-distance dependencies.First,a backbone feature extraction network DP-Transformer is proposed based on the hybrid structure of CNN and Transformer to learn the local and global relationship of defect features.Second,a multi-scale fusion transformer and a spatial crossover transformer are proposed,on which the feature decoder is constructed.In the decoder,the multi-scale fusion transformer provides valuable spatial information for the spatial cross transformer,so that the feature maps of different levels can be refined by the spatial cross transformer.The experimental results demonstrate that the proposed method can achieve a segmentation accuracy of 85.56% and a classification accuracy of 98.57%,with good defect geometric shape representation ability.Moreover,the arerage detection time reaches 0.0189 seconds.The lightweight design of proposed network provides theoretical support for subsequent deployment in practical applications.