A Research On The Automatic Detection Of Fabric Defects Based On Hybrid Feature Vector And One-Class Classification Detector

This dissertation aims at researching and developing practical computer-vision-based algorithms for automatic detection of woven fabric defects. Substituting computer vision for human eyes in the practice of fabric defect detection enables dramatic reduction of missing rate, increasing of production efficiency,guarantee of stability of the detection results, reform of defects information management and improvement of products quality monitoring and controlling. To make research on the algorithms of automatic detection of fabric defects is the key to realizing automation of fabric defects detection and the foundation of developing automatic cloth inspection machine as well. Therefore, rich interests in detection algorithms research have been aroused among the scientific researchers in the associated areas such as textile discipline, computer science and technology discipline and automation discipline both at home and abroad, making this research to be one of the studying focuses and difficulties of the frontier of textile discipline nowadays; at the same time, this research has also been representing one of the typical and important aspects of reforming and upgrading the traditional textile industry by advanced science and technology currently. On this account, this subject research is of good meaning in the sense of both theoretical and practical.A practical automatic fabric defects detection algorithm should meet the requirements of detection accuracy, one-class classification capability and real-time implementing by and large. Detection accuracy requirement demands that the detection false alarm rate and missing rate of the algorithm can be controlled at an acceptable low level simultaneously for the vast majority kinds of fabric defects. To satisfy this accuracy requirement, the algorithm should firstly ensure strong universality of the extracted feature vector for most kinds of detects, and secondly ensure high elaboration and strong feature vector data mining capability of the detector. One-class classification requirement demands that the designed detector must be of one-class classification purpose, namely, training of the detector can rely on the normal samples alone and not on both of the normal ones and the defective ones. Because of the great variety kind, form, shape and size, etc., of the fabric defects and the vast diversity of fabric texture and structure, training of the detector with defective samples cannot be implemented in practice, as a result, computer-vision-based automatic fabric defects detection is in nature a typical one-class classification task that takes texture image as its analyzing object. Real-time implementing requirement demands that data preprocessing, features extraction and detector’s discrimination can all be implemented fast enough to ensure the real-time detection. To make algorithms meet the above-mentioned requirements as far as possible, during the algorithms researching and developing stage, author of this dissertation presents three effective fabric defects detection algorithms that have not yet been reported so far in the research area of fabric defects detection in terms of whether extraction of feature vector or designing of one-class classification detector through theoretical exploration and experimental validation.This dissertation is divided into seven chapters.Chapter 1 is the general introduction, introducing the academic and industrial background of the research subject of this dissertation, analyzing the existing and concerned research achievements and their deficiencies made by the previous researchers of the research group, which majors in the applications of computer vision in textile process and where the author is currently with, and making clear the research objectives and content arrangement of the dissertation.Chapter 2 is a summarization of the existing literatures, analyzing and discussing the important literatures at home and abroad of the recent fifteen years on the basis of a brief introduction to the concerned concepts such as pattern recognition and texture analysis, etc., laying special emphasis on the discussion of the most related existing algorithms and pointing out values and limitations of the existing research achievements.Chapter 3 is an introduction of experimental samples collection and concerned data preprocessing technologies, all of which are the involved common problems of Chapter 4, Chapter 5 and Chapter 6. Specifically speaking, this chapter introduces the collection of defective fabric samples, category of fabric defects and the corresponding image demonstrations, acquiring and preprocessing of fabric texture images, determination of the sub-window size of fabric texture images, effectiveness verification of the texture image features, features normalization, samples datasets and samples distribution, and the indices for evaluation of the detection results of the algorithms, etc. Chapter 4 is the introduction of the first detection algorithm designed in this dissertation, which is mainly based on feature vector extracted by means of time series Autoregressive (AR) spectral analysis and based on a unilateral distance detector. This algorithm makes full use of the characteristics of periodicity and warpwise/fillingwise orientation of the fabric texture and takes the fact that most of the defects are warpwise or fillingwise into consideration, extracting feature vector from the one-dimensional time series obtained by fabric image projecting based on low order Burg AR spectral estimation method and detecting fabric defects with the self-designed unilateral distance detector. Because of its low computational complexity, this algorithm can meet the real-time implementing requirement. Detection results show that the general average false alarm rate and missing rate are 6.71% and 13.53%, respectively, indicating that this algorithm can also meet the detection accuracy requirement in the main.Chapter 5 is the introduction of the second detection algorithm designed in this dissertation, which is mainly based on a hybrid feature vector composed of multiple time series fractal features and a fuzzy-c-mean-clustering-based one-class classification detector. To describe the defective texture more finely and hence reduce detection errors, author of this dissertation proposes an original feature vector extraction concept that stresses the compromise of texture description ability of the features in the hybrid feature vector between general and detailed texture information, and between warpwise and fillingwise information as well. As a concrete realization of this concept, a novel multiple fractal feature vector is formed on the basis of the basic box dimension via the warpwise/fillingwize orientation characteristic of fabric texture and the basic texture period parameters. This combined multiple fractal feature vector is composed of one general fractal feature and four detailed fractal features, all of which are extracted on the basis of one-dimensional projecting series. In order to eliminate the disadvantage of possible unbalanced distribution of the original normal samples used for training and improve the ability of the detector in information mining of training data, a detector capable of one-class classification is self designed based on Fuzzy c-Mean (FCM) clustering algorithm, and optimizing of the parameters of the detector is also discussed. Experimental results indicate that, compared with that of the first algorithm, detection accuracy of this one has been improved to a large extent in that the general average false alarm rate and the missing rate of the second algorithm are 4.93% and 5.06%, respectively. However, extraction of the four detailed fractal features involves comparatively large computational complexity especially when dealing with fabric texture with high density. Therefore, this algorithm cannot meet the real-time implementing requirement for the moment and this is its major deficiency.Chapter 6 is the introduction of the third detection algorithm designed in this dissertation, which is mainly based on a hybrid feature vector composed of one fractal general feature and four Sobel-filtering-based detailed features and the support vector data description (SVDD) one-class classifier. A more robust and practical method for optimizing the detector parameters is also proposed. This algorithm borrows ideas from the second algorithm in the sense of taking into consideration of both general and detailed information, and both warpwise and fillingwize information in feature vector extraction. Difference of this algorithm from the second one is that another kind of detailed feature other than fractal is employed, i.e., Shannon Entropy detailed feature based on Sobel filtering is extracted. Detailed features’extraction speed of this algorithm is much faster than that of the second one, while at the same time, this algorithm preserves the advantage of the combined feature vector extraction concept mentioned above. Besides, the obvious complementarity between Sobel-filtering-based feature and fractal feature enables the hybrid feature vector of the third algorithm to be more reasonable in describing of texture information. As to detector, this algorithm adopts an advanced kernel machine learning methods, namely, the SVDD. SVDD is a special kind of one-class support vector machine, capable of effectively characterizing the distribution of normal training data and especially suitable for novelty detection. Experimental results of this algorithm are that the general average false alarm rate and missing rate are 4.61% and 4.09%, respectively. This algorithm is a little better than the second one as far as detection accuracy is concerned, and possesses the advantage of the first algorithm in detection speed. Therefore, this algorithm can be said to be the best algorithm among the three ones.Chapter 7 is the conclusions and outlook of the dissertation. Advantages and disadvantages of the three algorithms are briefly summarized; the achievements and deficiencies of the dissertation are also discussed; finally, the future research direction and emphases are put forward.The three algorithms of this dissertation are hardly more than three comprehensive and extrinsic materializations of the dissertation’s essential research achievements, which can be boiled down mainly to the following statements:1) Feature vector extraction concept taking into consideration of both general and detailed texture informationIn consideration of the diversity of fabric defects size and appearance and the insufficiency of one single feature in describing texture, and taking notice of the facts that fabric texture is of vertical and horizontal orientation and of obvious periodicity, this dissertation suggests that the extracted multiple features or feature vector should be capable of describing both general and detailed texture information so as to realize maximum complementation among different features and therefore more comprehensive characterization of the texture information. The multiple fractal feature vector of the second algorithm and feature vector composed of one fractal general feature and four sobel-filtering detailed features of the third algorithm are no other than reflections of this concept, which differs greatly from others who pay attentions only to unilateral function of the features.2) Detector design idea taking into consideration of both elaboration and one-class classification functionThe characteristics of fabric defects’diversity and being hard to completely collect determines that a practical detector must be capable of mining data information in depth and possess one-class classification function. One-class classifiers appearing in the existing literatures are mainly based on Euclidean distance or some simple thresholding methods, which are too rough to profoundly mine more implied information from the feature vector; while the more elaborate and advanced detectors such as neural network or support vector machine involved in existing literatures are devoid of one-class classification function in that training of these detector have to resort to a large quantity of defective samples. In view of this, we suggest that when designing a fabric defect detector, it is necessary that the detector is provided with functions of both one-class classification and in-depth information mining. One-class classification detector based on fuzzy c-mean algorithm and the support vector data description one-class classifier constructed in Chapter 5 and Chapter 6, respectively, are exactly the embodiment of the design idea, besides, neither application precedents or relevant reports of the two detectors can be retrieved in the filed of fabric defects detection.3) Real-time implementing conceive existing all along the whole detection algorithm procedureThis dissertation not only pursues the high detection accuracy goal of the detection algorithm but also emphasizes the real-time implementing requirement in the whole detection algorithm procedure. For this reason, in many links of the algorithm design the real-time implementing requirement has been taken into account. For example, a fast preprocessing algorithm for eliminating uneven lighting effect and enhancing image contrast is proposed. For another example, by making use of the fact that woven fabric textures and most of the fabric defects are of horizontal and/or vertical orientation, a projection operation is implemented which prompts features extraction to be carried out on the basis of one-dimensional time series rather than two-dimensional image and accordingly decreases the computational complexity dramatically. AR spectral features and all fractal-based features are extracted from one-dimensional time series. For a third example, estimation of AR spectral is based on low-order Burg algorithm characterized by recursion, which makes the involved amount of computation to be very small. For a fourth example, to preserve the superiority of the feature vector of the second algorithm in intensive description of textures and at the same time reduce computational complexity, the. faster Sobel-filtering-based Shannon entropy detailed features are employed to take the place of the original fractal-based detailed features in the design of the third algorithm. For a fifth example, we suggest that a larger weight index of the fuzzy c-mean clustering should be used in the case of sample quantity is relatively large so that the involved iteration number can be decreased substantially, etc.4) Sub-window partitioning method based on the periodicity of fabric textureA method of image sub-window partitioning based on the periodicity of fabric texture is proposed, which strengthens the stability of features and hence helps to reduce detection errors.