Woven Fabric Defect Detection Based On Non-negative Dictionary Learning

Fabric defect detection is a very important process in the production of the textile, whichis the key to ensuring product quality and improving the level of quality control. Traditional defectdetection mainly relies on artificial, which have many disadvantages, such as high missing rate,low productivity, without modern people-centered concept of the production and so on. It is urgentto carry out information technology and automation in the entire fabric defect detection industry.In recent years, expressing signal using the dictionary learning has received widespreadattention, especially in the field of computer vision. We propose a woven fabric defect detectionalgorithms which based on a non-negative dictionary learning of woven fabric defect detection.Compared to the traditional fixed or pre-defined dictionary expression method, the dictionarythrough studying is more able to adapt to the characteristics of signal and more effective on theexpression of input signal.This paper discusses the dictionary learning algorithm from the new idea of the wovenfabric texture on the spatial domain approximate expression. Firstly, the approximate expressionof fabric image is executed by non-negative dictionary learning method.Then the fabric defectdetection is projected based on this. The idea of detection has good practice application.For anintuitive understanding of the testing results, this paper finally compared the detection effect of thetwo main non-negative dictionary learning method.The following main topics of this research willbe described.(1) The method for obtaining the fabric image samplesDefect detection method is carried out based on the sub window, and each child windowis connected by column, then expanded as a column vector through head and tail of each column.Without loss of energy in the image information of the fabric and greatly reduce the amount ofcomputation.(2) The best dictionary elements obtained by using non-negative dictionary learningmethod of the test samples to reconstructThrough the method of non-negative dictionary learning based on active set, texture features only reflect the normal texture features from the normal sample, then the detectionsamples are reconstructed using the non-negative dictionaries. By comparing reconstruction errorbetween approximate image and the original image and determine whether the defect region.(3) Preferably the size of window and the size of dictionaryThe window size has a great influence on the effects of the defect detection, so thewindow size must be optimized. Normally we should consider the size and proportion of defectsin the sub window when selecting the size of the child window.The number of dictionary elementsdetermines the reconstruction error K. The more the number of dictionary elements, thereconstruction error is smaller, it has a great influence on the detection results. Throughcontinuous exploration in the course of the experiment, it hopes to find a sutitable dictionary, inthis case, good texture area will be reconstructed well and the defective part will be reconstructedwith larger reconstruction error.In this paper, the dictionary number of plain fabric is8,twill fabricis4and the size of sub window is32x32pixels.(4) Defect detection experiment and effectAfter the6528samples’ experiment, the results show that the proposed non-negativedictionary learning algorithm can make the error rate of less than10%, the detection rate greaterthan90%.(5) Comparing the effect of detection algorithm using two kinds of non-negativedictionary algorithmsBy comparing the detection effect from aspects from the false detection rate, falsenegative rate, image reconstruction and so on, it seems that the algorithm adopted in this paperhave a more intuitive understanding effect.After the6528samples’ experiment, it was tested that plain fabric false detection ratewas2.11%,detection rate was90.21%,twill fabric false detection rate was1.83%, the positive ratewas91.41%,which based on classical multiplicative rule; and the active set of plain fabric falsedetection rate is1.97%%,positive rate is91.79%%, twill fabric error detection rate is1.68%%,the detection rate is92.24%. Results showed that nonnegative dictionaries based on the active setis superior to that based on the multiplicative rule.In addition, real-time based on the active set issuperior to that based on the multiplicative rule.