Research On Tire Defect Detection Method Based On Deep Convolutional Network

Tire defect detection is the process of detecting and segmenting defect areas by applying manual observation or machine vision algorithms on tire X-ray images.The manual detection method is susceptible to visual fatigue,and has the disadvantages of low efficiency and high subjectivity.In contrast,the automatic detection technology can ensure the continuity of the production line,and has become an important and effective tool for improving manufacturing efficiency in industrial production.The performance and efficiency of detection algorithms directly affect product quality and production speed.Therefore,detection algorithms play an important role in the quality inspection of the tire industry.Existing automatic algorithms for tire defect detection mainly focus on distinguishing the difference between the defective area and the background(non-defective area).The key challenge of these methods is how to effectively extract features and accurately represent images.Constrained by the X-ray imaging equipment and the production environment,tire X-ray images generally contain noise and have low contrast,which poses challenges to automatic defect detection.Meanwhile,due to the similarity of the background texture as well as the diversity of defect textures and types in the tire image,it is difficult to effectively represent images based on hand-crafted feature descriptors.In addition,due to the imbalanced samples of defective and non-defective areas in the image and the requirements of real-time detection in industrial applications,tire defect detection is still a challenging task.To solve the challenges of detecting defects,we combine deep convolutional networks and propose two network models: a tire defect detection model based on a full convolutional neural network(FCN-Tire)and a multi-scale defect detection network(MDDN).Relying on the powerful self-learning and representation capabilities of neural networks,the FCN-Tire model with VGG as the basic structure can retain spatial feature information.The image features are first extracted by five stacked convolutional and pooling layers.As the number of layers increases,the more abstract the features obtained by the network.Therefore,the most abstract semantic features are mined and retained in feature maps derived from the last convolution layer.Then,feature maps derived from different depths are enlarged to the same size as the input image by up-sampling layers.To obtain the final detection result,these multi-scale maps are fused and the pixel-wise class prediction is obtained through the softmax layer.Compared with the traditional methods with handcrafted feature extraction operators,FCN-Tire can automatically learn tire feature representations to achieve dense prediction and pixel-wise segmentation.Experimental results show that FCN-Tire can obtain better detection results than traditional methods,and also demonstrate the feasibility of applying the deep convolution network to handle the tire defect detection task.Although FCN-Tire has obtained comparable detection results,extensive experiments have shown that FCN-Tire is not sensitive to defect small-scale defects.The primary cause is that some defective areas are ignored by pooling layers as backgrounds.To solve this issue,we develop a semantic-aware network on the basis of retaining the capability of FCN-Tire to extract deep abstract semantics.Meanwhile,a texture-aware network containing only four convolutional layers is proposed to supplement the detailed information filtered during extracting high-level semantics.Without pooling layers,detailed features can be fully retained.Finally,an ensemble learning structure is used to fuse multi-scale features to achieve the accurate defect detection.Compared with FCN-Tire,the ablation experimental results show that MDDN can achieve a detection accuracy gain by 4% without significantly increasing the computational complexity.In a word,based on the detailed analysis of tire image characteristics,this work explores the feasibility of applying deep convolutional networks to detect tire defects.By constructing two detection models,key issues such as texture insensitivity and data imbalance in the dataset are addressed.Experimental results show that proposed methods can obtain more accurate performance than most existing tire defect detection algorithms.