Research Of Defect Detection For Yarn-dyed Fabric Based On Image Analysis

With the improvement of life quality, people pay more attention to the clothing aestheticin daily life, thus the yarn-dyed fabric is favored by many consumers for its generous andsolemn style. Compared with grey fabric, diversified and small-batch production isemphasized for yarn-dyed fabric, and the color information of yarn is more emphasizedduring yarn-dyed fabric inspection, thus the defect detection for yarn-dyed fabric is moreimportant. At present, there are more reports on fabric defect detection based on machinevision, but the focus of the research is still grey fabric, and the related research of defectdetection for yarn-dyed is less. Therefore, algorithms of yarn-dyed fabric defect detection areresearched in this paper. The basic design idea of the detection algorithms is to selectdefective image as a template and to determine the fabric with defects or not throughcomparing the similarity of image being detected and template image.In order to fully master the basic experience of fabric defect detection and recognition,first, with grey fabrics as the researched objects, this paper puts forward a detection algorithmbased on characteristic spectrum of Fourier transform and correlation coefficient. Fabricimages can be transformed into frequency domain from spatial domain by FFT. Thecorresponding relation between fabric structure parameters and peaks in the spectrum isanalyzed and five characteristic parameters of image gray and texture are extracted.Correlation coefficient is used to represent the level of similarity between textures of imagesfor distinguishing defect-free fabric and defective fabric. Experiments show that the algorithmcan achieve accurate detection of six types of common defects, such as weft crackiness,cracked ends, stretched warp, looped weft, holes, and discover that the detection window hasan important effect on recognition results. The window size used in this paper is32pixels×32pixels.A method of pattern period detection for yarn-dyed fabric based on autocorrelationfunction is proposed in this paper. Defect-free image of yarn-dyed fabric is selected tocalculate its warp-directional and weft-directional autocorrelation curves respectively. Withappropriate window width, LOWESS algorithm is used to smooth the original curves. Peaksare located by finding local extremum points and pattern period is determined according tothe spacing of adjacent peaks. Among them, the width of smooth window is the key of thismethod. It is found that the width is related to image resolution and fabric density.An algorithm of defect detection for yarn-dyed fabric based on gray-level co-occurrencematrix (GLCM) and Euclidean distance is proposed in this paper. Compared with defect-freefabric, the element distribution of GLCM is changed when a defect appears; therefore, thefeature parameters such as contrast, energy, entropy, correlation, etc vary with it. Fabrictexture feature parameters affected by the distance and orientation are mainly discussed.When the algorithm is designed, feature parameters of GLCM aren’t extracted directly, instead,GLCM is taken as a characteristic matrix. Euclidean distance is used to represent the level ofsimilarity between GLCM of image being detected and template image to recognize defects.Various defects of different yarn-dyed fabric are tested, and the results show that thealgorithm has wide adaptability and high accuracy. An algorithm of defect detection for yarn-dyed fabric based on similarity measurementof projection sequence is also presented in this paper. First, fabric images are projected toobtain line and rank projection sequences. Secondly, use wavelet basis function db1andMallat algorithm to conduct3layers wavelet decomposition to projection sequences. Themean and standard deviation of each low-frequency coefficient are taken as characteristicparameters, and the feature vector consists of12sample data. Thirdly, the weighted Euclideandistance is used to measure the similarity of projection sequence between images beingdetected and template image to recognize defects. Experiment results show that the algorithmhas better detection effect for linear defects and block defects.