Research On Tire Defect Detection Algorithm Based On Deep Learning

Tire quality plays an important role in ensuring the stability of vehicle driving,and it is of great significance to detect the defects of tires quickly and accurately before leaving the factory.At present,the commonly used tire defect detection algorithm is based on the X-ray imaging technology to realize the tire defect recognition.However,due to the inherent structural shading of the tire on the X-ray imaging image of the tire,the background shading and defect position texture are easy to overlap,so it is difficult to judge the existence of defects by using the traditional image recognition method,which is prone to false detection and missed detection Aiming at the above problems,this paper proposes a tire defect detection algorithm based on deep learning(1)The research status of tire defect detection technology and target detection technology at home and abroad was described.Aiming at the main difficulties of tire X-ray image defect detection technology,a tire defect detection algorithm based on Faster RCNN was proposed;(2)The X-ray defect images of tires obtained from tire manufacturers were denoised.The images were marked by the calibration software developed by the laboratory,and the Pascal Voc2007 format tire defect data set was self-made,which provided a reliable data base for improving the accuracy of tire defect detection;(3)By analyzing the shortcomings of the original Faster RCNN algorithm in tire defect detection,the original algorithm is improved.The improved resnet101 is used to extract the features of tire defect image,and the candidate frame coordinates generated by RPN are mapped to the feature map by using ROI Align pooling method,which avoids the detection error caused by the twice quantization process of ROI Pooling method;(4)This paper introduces the feature weight heavy calibration module to the improved algorithm,compares and analyzes the effect of the feature weight heavy calibration module in different positions of the network through experiments,builds a feature extraction network based on SE net,makes full use of the global information of each feature channel,recalibrates the original features in channel dimension,and selectively enhances the channel with stronger feature representation ability.In addition,feature fusion module is introduced into the improved algorithm to realize the feature fusion of semantic information of deep feature map and detail information of shallow feature map.The network performance of the improved algorithm is further improved;(5)The experimental results show that the average accuracy of the proposed method in thin line,impurity and bubble defects reaches 95.6%,97.1% and 92.5%,respectively.Compared with the improved network strict judgment rate and missing detection rate,the proposed method is reasonable and effective.