Tyre Defect Detection By X-ray Image Based On Deep Learning

With the full boosting of "Made in China 2025",Automation,Informatization,Internet of Things,and Digitalization of the tyre industry is an inevitable trend.Tyre quality inspection is an extremely important process in tyre production.The realization of automatic detection of defects based on tyre X-ray images has become an important research topic.Aiming at the situation and shortcomings of manual quality inspection in tyre production,combined with obeject detection algorithms in the field of computer vision,this paper conducts research on tyre X-ray image defect detection technology based on deep learning to provide a feasible automated defect detection model for tyre quality inspection.This thesis uses deep learning technology to implement an end-to-end automatic detection model for tyre X-ray image defect detection.This thesis at first applies Faster R-CNN object detection algorithm to tyre quality inspection to verify the feasibility and effectiveness of automatic tyre defect detection.Further,aiming at the characteristics of the tyre X-ray image and the shortcomings in the Faster R-CNN network structure,innovative algorithms were proposed in the image preprocessing stage,model training stage and model testing stage to improve the applicability and robustness of the model.The main research contents of this thesis are as follows:Firstly,aiming at the tyre quality inspection process,a defect detection model based on Hybrid Faster R-CNN is proposed.The original image and different preprocessed images of the same tyre are input to multiple detection models for training respectively.Preprocessing methods include histogram equalization,contrast-limited histogram equalization,and Laplace transform.The outputs of each model are combined by the non-maximum suppression algorithm as the final detection result of Hybrid Faster R-CNN.Experimental results show that mixing multiple models improves the detection performance of the model.The mAP of the Hybrid Faster R-CNN raises up to 53.6%,which exceeds Faster R-CNN model by 1.3%.Secondly,in order to solve the poor performance on small-scale and multi-scale defects caused by using only single-scale feature maps in Faster R-CNN model,this thesis designs DCMS-FRCNN(Deformable-Convolution Multi-Scale Faster R-CNN)model.The ResNet-101 feature extraction network was reconstructed using the deformable convolution module,and the deformable pooling module was integrated into the RoI pooling network.Meanwhile the multi-scale RPN structure is used to make full use of deep and shallow layers and feature maps of different scales in the feature extract network,which further enhanced the invariance of the model to the deformation of tyre defects.The experimental results show that the mAP of the DCMS-FRCNN model reaches 61.1%.Owing to that the defects and the background in tyre X-ray image are essentially texture characteristics,the similarity between the two can reflect the characteristics of the defects to a certain extent.Finally this thesis proposes an improved detection algorithm based on background feature information.When detecting a defective image,the background features of the corresponding position in the normal image are extracted in parallel,the cosine similarity between the defective feature vector and the background feature vector is calculated,and used to re-rank the probability of the region proposals.Applying this testing algorithm to the former Faster R-CNN model,the mAP of the detection model was further improved to 62.7%.