| Tires represent the unique medium of contact between an automobile and the ground,and their defect detection plays a critical role in determining the factory quality of tires,thus affecting driving safety.In the era of Industry 4.0,automating,digitizing,and developing intelligent tire defect detection methods are vital requirements for the successful implementation of China’s "Made in China 2025" strategy.Currently,the majority of small and medium-sized tire manufacturers in China still rely on manual visual inspection,which entails high labor intensity and the potential for misjudgment.Furthermore,foreign-imported equipment poses challenges with varying tire models and limited interregional communication.Thus,designing automated tire detection techniques has immense significance.Deep learning technology has emerged as a mainstream technology of artificial intelligence,particularly in the field of object detection.As such,research into intelligent non-destructive testing of tires based on deep learning represents a vital research topic.In this study,I aim to investigate the intelligent non-destructive testing of tires using deep learning,including the detection of bubbles in tire crown laser speckle images and root opening,sparse line,and bubble defect detection in tire X-ray images,and propose corresponding solutions.The main contributions of this paper are as follows:(1)Two datasets of tire defect images were established,laying the foundation for subsequent research;(2)For the bubble defect detection in tire crown laser speckle images,a multi-level and multi-directional feature pyramid network(TYFPN)was proposed under the Faster R-CNN detection framework to address the detection difficulties such as small objects,small differences between objects and backgrounds,and internal differences within the objects.The effectiveness of TYFPN was demonstrated through k-fold cross-validation experiments and ablation experiments.Experimental results show that TYFPN has improved the m AP@[0.5: 0.95] and m AP@0.5 indicators by 2.08% and 2.4%,respectively,with better detection performance and lower miss detection rate.(3)For the research on tire X-ray image defect detection,it has been found during the research process that there is a high correlation between the object and background,and the determination of the object requires the combination of background information.A comparative experiment was designed to select the best-performing object detection architecture and backbone network.Inspired by the experimental process and results,the neck network was redesigned,and a DWFPN based on Depth-wise convolution was proposed.It has been experimentally proven that the m AP@0.5 evaluation index of DWFPN has increased by 1.86%,the parameter quantity has decreased by 2.3M,5.1%,the computation has decreased by 49.09 G,23.5% and the model size has decreased by26 MB,5.03%.Therefore,DWFPN is more suitable for tire X-ray defect detection.For detection scenarios where the object is highly correlated with the background,this paper concludes that expanding the effective perceptual field can be effective in its detection.(4)Deployed the network model for tire X-ray defect detection and compared the speed and accuracy under different quantization methods,selecting the appropriate quantization method based on specific scene requirements.Furthermore,compared the miss rate and false alarm rate under different network model confidence score thresholds in simulated production environments.Based on the principle of minimizing the miss rate in tire production,selected a confidence score threshold of 0.4,with miss rates of2.14% for Bubbles,0.41% for Rootbreak defects,and 3.14% for Sparselines.Reasonable detection schemes have been proposed for two types of tire defect detection scenarios,and detection ideas have also been provided for scenarios with small object detection and high correlation between objects and backgrounds,achieving good research results in tire defect detection and making certain contributions to the research of tire intelligent detection field. |
