Research On Tire Defect Detection Method Based On Object Detection

Tires play the role of load bearing and shock absorption in the vehicle structure,and their quality directly affects driving safety.In the tire production process,defect detection is a key link to ensure product quality,and tire defect detection methods are of great significance to traffic safety and the development of the tire industry.Traditional tire defect detection mainly relies on X-ray imaging technology to collect tire X-ray images,and then uses artificial naked eye detection or machine learning-based methods for detection.Artificial naked eye detection of tire defects has subjectivity and the possibility of misjudgment.In contrast,machine learning methods have higher accuracy and stability through training and learning with a large amount of data,but manual design of features is required,which is time-consuming,laborious and The problem of over-reliance on prior knowledge.In addition,the diversity of tire defects,large scale span,low contrast,anisotropic texture background and other characteristics make tire defect detection based on traditional methods still face challenges.Under the background of upgrading from traditional manufacturing to intelligent manufacturing,the intelligent detection method of tire defects has become a hot issue in the field of defect detection.Object detection algorithms based on deep learning have high precision,robustness,scalability,and real-time performance,so they have become popular methods in research.This paper takes the X-ray images of passenger car radial tires(hereinafter collectively referred to as tires)as the research object,and proposes an attention module,a feature pyramid module and a feature enhancement module to solve the problem of how to improve the detection accuracy and detection speed,and based on the above modules and the YOLO series Network Architecture Designs two tire defect detection models based on object detection,and performs experimental verification of the proposed method.The main research contents and contributions of the paper are as follows:(1)Propose a tire defect detection method based on improved YOLOv4.Based on the YOLOv4 algorithm,three improvements are proposed to the network according to the characteristics of defects in tire X-ray images.First of all,aiming at the problem that the anchor frame size obtained by clustering on the MS COCO data set cannot be well applied to tire defect detection,the K-means++clustering algorithm is used to optimize the anchor frame size of the tire data set,which improves the tire defect detection.Accuracy of defect detection.Secondly,aiming at the problem of low detection accuracy caused by the imbalance of positive and negative samples in the network training process,based on the Focal Loss idea,the confidence loss function is introduced to achieve a certain regularization effect.Finally,in view of the difficulty of target detection caused by the large scale span of tire defects,by introducing the Adaptively Spatial Feature Fusion(ASFF)module into the feature fusion network of YOLOv4,adaptively adjust the feature layers of each scale when fused.The spatial weights of the network enhance the representational power of the network and the ability to process multi-scale features.(2)A lightweight tire defect detection method based on attention mechanism is proposed.In order to achieve industrial online detection and balance detection accuracy and efficiency,the Convolutional Block Attention Module(CBAM)was introduced,and the Multi-scale Self-attention Feature Enhancement Module(Multi-scale Self-attention Feature Enhancement Module,MSAM)and Feature Pyramid Networks(FPN)structure.First of all,for low-contrast defects and complex texture backgrounds in tires,CBAM is introduced into the feature fusion network of YOLOv4-tiny to enhance the expression of useful features for tire defects and weaken irrelevant features such as tire texture backgrounds.Secondly,aiming at the problem that the model cannot capture global context information,MSAM is proposed to expand the receptive field,capture long-distance dependencies at the same time,and model global context features.Finally,in view of the problem that multi-scale feature information cannot be fully utilized and the detection of small target defects is difficult,a new feature pyramid network FPN structure guided by MSAM and CBAM is proposed,and the feature fusion network of YOLOv4-tiny is improved.By strengthening the low-level and high-level The fusion of feature information improves the representation ability of multi-scale targets.