Fabric Defect Detection Method Based On Deep Learning

The textile industry occupies a very important position in the modern national economy,and fabric defect detection is an important link in fabric quality control.But due to the variety of fabric defects,it brings a challenge to the defect detection.In recent years,the application of deep learning in various fields of images has achieved great success,providing a new method for efficient detection of fabric defects.It is well known that deep learning models not only need a large number of training samples to update parameters,but also sometimes need manual annotation,which is time-consuming and laborious.Fabric defects,as an abnormal phenomenon,can not be collected a large number of data.Under the condition of insufficient data,it is difficult for the network to learn complex texture and all kinds of defect information effectively,and therefore can not achieve satisfactory accuracy.In view of the above questions,this paper has carried on the related research and the experiment,the main content is as follows:(1)Aiming at the problem that the network does not perform well on various defects of complex textures in the case of a small amount of data,the paper proposes a semantic segmentation network to complete the fabric defects detection.Semantic segmentation networks usually require a large number of paired samples for training,while the existing labeled fabric data sets are limited,and manual labeling requires a lot of time and effort.Therefore,the paper proposes a defect sample-pair generation module,it doesn't have to be manually labeled,only need to select 5 normal samples of the same fabric and send them to the module,and 3000 pairs of training samples will be generated.Then the generated training samples are sent to the U-Net network that incorporates the attention mechanism for supervised training.Experimental results show that the algorithm performs well on various defects of various fabrics,with an average accuracy rate of 98%.(2)Aiming at the problem of insufficient defect samples,an unsupervised defect detection method that only needs normal samples for training is introduced.This method introduces an inverter on the basis of the traditional deep convolution generative adversarial networks.With the help of this inverter,the network can reconstruct the input image,and the discriminator to make the reconstructed image more realistic.In addition,in order to increase the robustness of the model,a converter T is added to the network,which will destroy the normal training samples to a certain extent,and send the damaged samples to the inverter and generator to reconstruction.In the test phase,because the network has learned the distribution of normal samples,if there are defects in the test image,the network will reconstruct the background,that is,reconstruct the defect area into normal texture,so that there will be differences in the defect area before and after reconstruction.Take the absolute difference between the original image and the reconstructed image will obtain a residual image that containing defect information.Then perform threshold segmentation on the residual image to obtain the final defect detection result image.(3)This chapter introduces in detail the construction of the experimental platform and the realization of the fabric defect detection system.It describes the construction of the PyTorch deep learning framework based on the Linux system,and uses tkinter to develop a fabric defect detection system based on the paper algorithm.On the system interface,different fabrics can be selected for defect detection,or generate random defects on selected images for detection and view the detection results.