Research On Automatic Detection For Yarn-dyed Fabric Defect Based On Machine Vision And Image Processing

Traditionally, fabric defect detection depends on manual work, which is easy to be affected by subjective factors in the detection process and has very low efficiency. With the development of automatic control and information technology, an automatic defect detection system gradually replaces manual testing, and has been an important means of controlling fabric quality. However, the present researches are mainly aimed at the automatic defect detection for gray fabric. Therefore, the paper studies automatic defect detection theory and algorithm based on machine vision and image processing, which is applied in broad and close yarn-dyed fabric. Many key problems are solved for detecting and classifying defects on the surface of yarn-dyed fabric with patterns. It is also the purpose of our research. The main content of the paper involves the research status of automatic detection system for fabric defect, design of detection and measure hardware platform, enhancement method of fabric textures based on fractional differentiation, intelligent detection algorithm of defects of yarn-dyed fabrics by energy-based local binary patterns, as well as yarn-dyed fabric defect characterization and classification method using combined features and support-vector-machine (SVM).First, the paper reviews the research progress about automatic detection system of fabric defects. This section firstly expounds our research’s significance, that is, application of automatic detection technology can greatly improve the labor productivity and corporate profits. The research background is then surveyed in the chapter. The characteristic and application for three commercial detection systems of fabric defects in the world are simply introduced, and the necessity of our research is concluded. According to different kinds of detecting objects, the paper lastly proposes automatic defect detection methods including gray fabric, gray pattern fabric and yarn-dyed fabric. The study on automatic detection algorithm for gray fabric is relatively mature. The statistic method in spatial domain divides a testing fabric image into differently statistical characteristic regions to segment defects; The method in frequency domain detects fabric defects by the similarity between the cyclical of basic texture primitive (organizational structure) and spectrum characteristic; Through mapping fabric textures into geometrical modeling, the model-based approaches can regard defect detection as hypothesis test for the model. The method of defect detection for gray pattern fabric which is researched less mainly comprises template-based image matching and repeating-unit-pattern-based window thresholding. The defect detection for yarn-dyed fabric needs to consider color model and more types of defects (weaving defects and color defects). There have been some algorithms for the defect detection of rigid materials, such as ceramic tile, wood, etc, but so far that of flexible materials, such as yarn-dyed fabric, print fabric, etc, has no a breakthrough.Second, the paper introduces the design scheme of hardware architecture for automatic defect detection system of yarn-dyed fabric. The overall design of the hardware architecture is firstly presented. This chapter then focuses on the hardware subsystem design for fabric image acquisition:LED lighting source is chosen based on the principle of illumination matching, and the forward and back lighting structure plan is discussed; When CCD cameras and image acquisition cards are selected, the selection factors which should be taken into consideration are detailed; Since image measurement of defect size is a new function in our research, the CCD camera subsystem needs to be calibrated beforehand. After image measuring principle and CCD camera calibration theory are introduced, the parameters of accurate spatial position and posture for each camera are calculated. According to the requirements of measuring precision for yarn-dyed fabric defects, pixel equivalent is determined through the experiments. Finally, the FPGA (field programmable logic array) interface of special development board is designed and the DSP (digital signal processor) is selected, which is the core of the whole hardware system. Since detection and recognition needs to deal with large amounts of data, the high performance DSP is specialized in the operation of image data. Thus, signal control and data transmission are implemented by multi-channel FPGA. The synchronous parallel image processing plan greatly improves the speed of the whole system, which reveals the feasibility and practicability of hardware system design.Third, the paper discusses the use of fractional differentiation to develop an effective enhancement method for fabric image with abundant textures, which belongs to the image pre-processing of automatic detection for fabric defects. The aim of image enhancement is to improve the visual quality of images, or to convert to a more suitable form in order to analyze and processing. The Grumwald-Letnikov (G-L) definition of fractional differentiation is in Euclidean space in the paper, and it is the results that differential order is extended from integral step-size to fractional step-size. Additionally, it should be known that there are a lot of textures in a fabric image. The fractional differentiation is a continuation of integer-order differentiation, thus similarly the fractional differential operator can also realize sharpening enhancement of image textures. By the analysis of amplitude-frequency characteristics, fractional differentiation can highlight the fabric image edges, improve the fabric textures and nonlinearly keep the details of the fabric image smooth areas; By the analysis of detection between stability coefficients and image textures in the dynamic theory, fractional differentiation not only nonlinearly enhances the contour features in the low-frequency area, but also highlights high-frequency edge features in those areas where gray changes remarkably; From the coefficients of fractional differential polynomial, isotropic operator is constructed to enhance fabric images. Through the respective experiments of defect images for gray fabric and yarn-dyed fabric, the effectiveness of enhancement for fabric texture image is proved by qualitative and quantitative methods. Based on the texture edge map of enhanced fabric image, the favorable enhancement is also confirmed indirectly by region homogeneity measure.Fourth, the paper designs and realizes the algorithm of intelligent defect detection of yarn-dyed fabrics by energy-based local binary patterns. Its purpose is to finish fast and effective detection of defects of yarn-dyed fabric via computer vision, and to consider two kinds of texture characteristics, that is, color and structure. The algorithm process is as follows:(1) The yarn-dyed fabric image enhanced by fractional differentiation is first converted from RGB true color space to L*a*b*color space.(2) In this color space, energy-based feature images are acquired by image fusion after the Log-Gabor filter filters chromatic and brightness component images. It solves the problem that color and structure defects can be appeared in the same energy feature image.(3) Through the analysis of energy-based feature images for yarn-dyed fabrics, the defects usually are locally brighter areas that have irregularity and non-uniformity and range from several pixels to dozens of pixels. The normal patterns in the background are regular and uniform. Therefore, with the help of the appropriate mechanism of scanning and comparing, what is needed is a local operator of the textures that could be used to detect defects in the energy-based feature images. The relations between energy and the local binary pattern are defined as a new concept called energy-based local binary patterns (ELBP), and the operator has the invariability for simple affine deformation.(4) Based on the minimum repeat pattern units in the defective reference images and testing images, many windows are segmented uniformly, and the ELBP feature vectors are obtained for each definition window in the energy-based feature images.(5) The defective threshold is found by likelihood estimation in the training stage. The defective windows are detected by comparing the threshold with detection windows of tested images in the detection stage, which can segment defect area. The proposed method can detect chromatic and structural defects. Experimental results for the defect detection from several collections of yarn-dyed fabrics indicate that a detection success rate of more than94.09%is achieved for the proposed method. The speed of test is also fast, and it is suitable for off-line testing.Finally, the paper explores a novel defect evaluation method that uses combined features and modified support vector machine (SVM) classifiers to characterize and classify the defects of yarn-dyed fabrics. This section firstly introduces the extraction of combination feature set. The geometrical features are defined, such as weft length, warp length, weft length to warp length ratio, perimeter, area, and roundness, and six geometrical parameters are extracted from the binary defect images. Concurrently, three textural parameters that characterize coarseness, contrast, and directionality can be defined and extracted on the basis of the textural energy defect images. The extraction processes of two types of feature parameters can combine into the extractor of combination features. The combination feature set which outputs in the extractor can quantitatively descript the surface characteristics of yarn-dyed fabric defects. These parameters are also used as the inputs of optimized SVM classifiers to obtain overall defect classes in accordance with the Chinese National Standard of Yarn-dyed Pattern Fabrics (GB/T22851-2009), that are cracked-ends (slack-ends, double-ends and mix-end), reedmark, broken picks (double-pick and looped-weft), weft-crackiness, hole (float, knot and gout), and stain (burl-mark, bruise and pan-color). This chapter then discusses the radial basis kernel SVM classifier and the optimization of general SVM classification model. Since the effectiveness of the SVM classification scheme largely depends on the selection of the classifier parameters, a ’grid searching’ procedure is used to determine the best selection of two parameters to achieve the highest classification accuracy in the application. When the combined feature set is used as the inputs of the SVM classifier, the LOOCV (leave-one-out cross-validation) is designed to avoid the biased classifications. The experiment samples are180fabric defect images for three types of yarn-dyed fabrics with different patterns. The cross-validation test on the yarn-dyed fabric defect classifications indicates that the defect categories of more than91%can be recognized correctly by using the SVM classification scheme. Compared with the accuracy of89.4%for BPNN, this algorithm is more robust and effective.