Detection And Classification Of Textile Defects Based On Computer Vision

In the textile industry,quality control is getting more and more important by enterprises and consumers.Among them,the defects in textile fabrics are the most common quality factor.Therefore,the detection of textile defects is also a key element in quality assurance.Currently,the traditional human eye detection method are still being used by some companies.This method requires a lot of manpower consumption,and detecting standards are not uniform,there is no guarantee of accuracy.Therefore,the computer vision and machine learning tools should be the trend to solve the problem of textile quality control.In this paper,according to the characteristics of common textiles,a method to detect the defects of textiles was proposed.This method can support the identification and location of defects,classify the defects,and provide the final judgment conclusion.The recognition of many kinds of fabrics is also supported.For the classification of fabric types,Gray Level Co-occurrence Matrix(GLCM)is used to extract feature values from cloth images and find the feature vector of each sample image.K Nearest Neighbor Algorithm is used to classify fabric types.The algorithm is based on the training set for classification.In this paper,272 sample images of three kinds of fabrics were taken,and the accuracy of classification prediction under different distance measurement methods,different voting weights,and different K values was tested.The final classification accuracy is 98.5%.Gabor filter and improved GLCM method were adopted to detect defects.Before the detection,a mask sharpening method and light-ratio balance method for image preprocessing was designed.The mask sharpening method ensures sharp details without noise due to sharpening.The light-ratio balance method can balance the light intensity of various parts of the image and reduce the error of subsequent analysis.Gabor filter is sensitive to the image texture.A 5-scale 8-direction Gabor filter bank was adopted to achieve the coverage of most of the textures.From the 40 filtered images,the best 10 filtered images were selected for fusion.The traditional GLCM analysis method based on small windows is computationally intensive and time consuming.In this paper,an improved method is proposed to analyze the entire image quickly with general accuracy and then recognize locally in high accuracy,which can significantly increase the recognition speed.Finally,the fusion results of the two methods achieved overall 94.33% defects detected rate.For the category classification,this paper adopts support vector machine,BP neural network,CART decision tree and fusion classification method respectively.Through experimental analysis,the best prediction parameters of each classifier were determined.The average prediction success rate of SVM was 90.6%,CART decision tree was 91.13%,and BP neural network was 94.47%.Different classifiers have better recognition accuracy for some other types of defects.Therefore,in the fusion classification method,this paper analyzes the classification accuracy of different classifiers for different types of defects,and adopts their classification results according to the classification of the classifiers.Finally,the sample's class classification accuracy rate of 3 kinds of fabrics reaches 98.95%,97.86% and 95.33%.Finally,a set of visual detection system based on MATLAB was introduced,which can give the conclusions of the quality detection of fabrics and suggestions for improvement,and supports batch detection and real-time correction functions.The experiment proves that the design can complete the detection and classification work of cloth defects.