Fabric Defect Detection Method Based On Machine Vision

The defect of fabric is the main factor that causes the degradation of fabric quality.As an important part of the process of textile manufacturing,defect detection mainly relies on manual work.This testing method is not only inefficient and easy to be affected by subjective factors,which becomes an obstacle to the intelligent development of the textile manufacturing industry.Based on this,this thesis proposes a defect detection method based on unblemished fabric image training,which solves the problem of textile feature recognition under the condition of small samples,and lightens the defect detection network model,greatly reducing the number of training parameters and the amount of calculation,while ensuring the detection accuracy,improving the detection speed and reducing the memory cost.The work of this thesis mainly includes the following four aspects:(1)The input image is preprocessed to improve the accuracy of flaw recognition in the whole network.According to the peak signal-to-noise ratio(PSNR),the common image preprocessing methods are compared,and the optimal processing method is selected.Experiments show that the image preprocessing method can improve the accuracy of fabric defect detection.(2)Feature extraction is carried out on the preprocessed image.There are often small regional defects in fabric images,which seriously affect the accuracy of network detection.YOLOv3 uses a multi-scale feature fusion network to integrate the medium and deep features of the feature extraction network Dark Net-53 with the shallow features to obtain richer feature information,so as to optimize the defect detection network.However,the method of YOLOv3 repeatedly predicting the output feature information affects the detection speed of the network,so an improved feature extraction method based on Dark Net-53 is proposed.The proposed method abandons the steps of multiple predictions of the output feature information by YOLOv3,and fuses the feature information of different levels for one output,which improves the detection speed of network model and the detection accuracy of small target defects.In the test set,the detection accuracy of small target defects reaches 91.5% and the detection speed reaches 49 FPS.(3)The feature extracted from the network is identified by flaw classification.In textile production,it is very difficult to obtain a sufficient number of defect samples,and labeling the data is a tedious task.Therefore,in this thesis,a network based on unblemished fabric image training is used for defect detection.After 100 normal fabric samples are collected and placed into the network training,defect detection is carried out on the test image.The method proposed in this thesis greatly reduces the workload of data collection and labeling.In addition,the output of the detection network is a reconstructed image,and the normal area of the test image is reconstructed into black,while the defective area is reconstructed into white.According to the experimental results,the identification rate of fabric defect detection method proposed in this thesis is up to 93.3%.(4)The lightweight of fabric defect detection network is further studied.In view of the application scenario that it is difficult to equip an efficient inspection device in the production site and deploy the inspection model on the mobile terminal device,this thesis proposes a lightweight network algorithm for fabric defect detection.In other words,to reduce and optimize the parameters of the whole network model,the classical deep separable convolution in Mobile Net is used to replace the ordinary convolution of 3×3 size in the network model.The number of network parameters and calculation amount is reduced by 8/9 after the lightweight,and the memory usage is reduced by more than half.