Research On Tire Defect Detection Technology Based On Deep Learning With Small Sample Set

Tire defect detection is a necessary link to ensure tire quality.At present,the research of scholars at home and abroad mainly focuses on the defects of large sample set tires,and how to detect the types of defects and determine the location of defects in the small sample set tires X-ray image is still a valuable research topic.In this paper,a deep convolution neural network model is built to improve the detection speed,detection accuracy and recall rate of tire X-ray image defects.The main research contents and innovative work are as follows:(1)Through the preliminary experiment of deep learning object detection,yolov3 network model is determined as the model framework of small sample set tire X-ray image defect detection.First of all,build a commonly used deep learning object detection model,and use small sample image data to carry out a preliminary experiment of deep learning object detection;finally,from the experimental results,we can see that yolov3 model is the most suitable model for small sample tire X-ray image defect detection,and analyze and infer the reasons of low detection accuracy and recall rate.(2)Aiming at the problem of insufficient image data(small sample set)in tire Xray image samples,the methods of data enhancement,image transformation and image cutting are proposed to process the sample data.First,two methods of data enhancement are used to realize the initial expansion of the data set;then,three methods of image transformation,namely Fourier transform,wavelet transform and histogram equalization,are used to process the image and expand the image data set of tire defect samples to four times of the original one;finally,the original defect image size is reduced by using the image cutting method based on the annotation frame,The defect can be separated effectively,and the purpose of sample image data amplification can be further realized.In addition,the image transformation method can also highlight the details of tire defects.(3)In view of the two problems of low detection accuracy,low recall rate and slow detection speed caused by the small proportion of tire defect area(small object),the yolov3 improved network model is proposed.The optimized network structure of darknet-39 is used for feature extraction,which effectively improves the detection speed of tire defects.By modifying the multi-scale detection module,the original threescale detection is modified to four scale detection,which further integrates the high-resolution shallow features and high-level features of high semantic information,and improves the detection effect of small object tire defects.(4)Through the contrast experiment,it is found that the image transformation method has the function of data enhancement,and emphasizes the image defect feature of small sample set,which effectively solves the problems of the defect feature is not obvious and the sample image data is insufficient.In addition,experiments show that yolov3 improved network model has a good detection effect on small object tire defects,which makes the detection accuracy rate reach 95.8%,the recall rate from 58.3% to 80.5%.