Automated Woven Fabric Defect Detection Using Dictionary Learning

Surface quality control of raw woven fabrics is a vital step for the modern textile industry as defects on the fabric surface can have a significant impact on the value and grading of the final product. At present, fabric inspection in textile mills is still performed manually by human inspectors, a process that suffers from both low efficiency and high labor intensity and cannot always meet the needs of high quality mass manufacturing. As one of important industrial applications, computer vision-based fabric defect inspection system can not only avoid some of the limitations of employing human inspectors, but can also provide an objective, stable and high efficient detection results of final product.This research mainly focuses on the development of useful defect detection algorithms for woven fabrics with computer vision. Considered as the core of an automated fabric defect inspection system, detection algorithm is to recognize defective pixels or areas in fabric images according to a certain decision rule that differentiates defective textures from normal. In general, detection algorithm involves extracting features and designing classifier. Feature extraction is to seek certain discriminative features that can succinctly describe defect-free regions (homogeneous texture) and thereby are sensitive to defects differentiated from the homogeneous background; classifier design aims to learn model parameters with respect to a specified classification problem, in order to make expected decision to the extracted features. However, such feature extraction based approach always confronts feature selection problem, since there is no straightforward way to guarantee the optimality of features used in case of absence of negative samples (defects) involving in training process.To devise detection algorithm more efficiently, a comprehensive analysis of existing work and essentials characteristics of fabric textures was conducted, which suggested that template matching is a quite ideal method for fabric detect detection, and a necessary condition of applying this method is to make an effective description or representation for defect-free fabric textures. Therefore, this thesis comes as a fairly new route for detection algorithm design. Firstly, dictionary learning is used to make an approximation representation for normal fabric textures; secondly, the template matching scheme is used to design detection algorithms based on the approximated results with the learned dictionary. Thus, the complex defect detection can be converted into a template matching problem, whereby the feature selection problem can be somehow alleviated and the performance of detection algorithm does not depend on the types of defect, improving the generality of designed algorithm. This thesis consists of eight chapters, whose organization is given as follows: In chapter one, general introduction concerning this thesis is made, including academic and industrial background, achievements and drawbacks of previous work conducted by research group, and research focus and contents of this research.In chapter two, an overall review of previous defect detection methods is provided. Firstly, the fundamental concepts of related disciplines are introduced; secondly, the existing methods are classified into two categories according to their detection schemes:feature extraction-based and non-feature extraction-based, and an overall survey of related work over twenty years is presented, in which the close related methods are addressed; finally, a summary on those aforementioned methods is concluded and commented.In chapter three, the major difficulties making fabric defect detection a challenging problem are drawn up after analyzing prior work and fabric textures. With this understanding, the idea of using indirectly template matching method for detecting defect is presented. Then, primary experiments with low-dimension approximation using principal component analysis are made, and a more general method of using dictionary learning for approximation representation of fabric images is introduced. Necessary experiments are also conducted to confirm the efficiency of dictionary learning method in approximating fabric images.In chapter four, an unsupervised detection algorithm using dictionary learning is investigated, and the way to select optimal dictionary size is discussed as well. The proposed method is based on unsupervised strategy, and the dictionary can be learned from the testing image straitforwardly. By selecting proper dictionary size, the normal fabric textures can be approximated well, and the defective textures cannot. With this, defective areas can be well enhanced in the residual image obtained from subtracting original image from approximated one, and then a simple threshold method can be used to segment defects. To adapt more defect types difficult to identify such as linear defect, an improved method using image rotation and linear projection are proposed. Experimental results on20datasets show the proposed method can achieve87.3%of CDR and8.8%of FAR. To further confirm the efficiency of the proposed method, a comparison with the Fourier method is conducted.In chapter five, a semi-supervised detection algorithm using dictionary learning is described. This algorithm is to identify fabric defects based on patches (sub-windows), but not attempt to extract features from the raw patches straightforwardly. Instead, this method firstly aims to approximate those raw patches with the learned dictionary, and then manages to extract more meaningful similarity features between the raw patches and their approximated fashions, and the advanced one-classifier, namely support vector data description, is applied to classify defects. The proposed method is similar to template matchin, but using adaptive template, not needing to consider template choosing and alignment problem. With this, defect detection can be conducted based on the more reasonable template matching, rendering the the performance of the proposed method does not depend on the specific defect types but depend on how the anomalous a defect makes. And a heuristic method to determine proper dictionary size is also discussed. Experimental results on20datasets show the proposed semi-supervised method can achieve95.5%of CDR under1.9%of FARIn chapter six, an extension of the semi-supervised detection algorithm in chapter five is presented. Aiming at its real-time implement issue, an analysis on computational complexity is investigated, suggesting that orthogonal dictionary is good choice for reducing computational complexity. Considering the future research, this chapter also extends the semi-supervised to non-negative dictionary and sparse dictionary. Experimental results on the three extensions show that sparse dictionary has the best performance, but orthogonal dictionary have the best overall performance both in accuracy and real-time implement.In chapter seven, the real-time experiment using orthogonal dictionary is conducted. To realize real-time, the real-time fabric defect detection platform devised by our research group is used, and the detailed configurations in terms of hardware and software are discussed and described, including hardware parameters calculating, real-time condition for a detection algorithm and realize procedures and so on. Experimental results with160m real world fabric using orthogonal dictionary algorithm demonstrates that this algorithm can achieve89%of CDR and3.9%of FAR in speed of30m/min.Finally, a summary and outlook of this thesis are given in chapter eight, where a concise summary and drawbacks of this research are presented and some recommendations for further research are suggested as well.Contributions of this research:(1) Starting from representing fabric texture in spatial domain, a new fabric defect detection framework with indirectly template matching is proposed. Such detection scheme is not only able to bypass the feature selection problem confronted by feature extraction based methods, but can convert the complex fabric defect detection into a template matching problem to solve as well. In addition, template matching scheme for defect detection is quite suited for the logic of computer in discriminating defects, and also a potential direction for future research.(2) Making full use of periodic property of fabric texture and the flexibility of dictionary learning in approximation representation, a new unsupervised detection algorithm based on dictionary learning is proposed. Aiming at the poor performance on linear defects, an improved strategy based on image rotation and linear projection is proposed. Meanwhile, the way to select optimal dictionary size using a cost function is designed. The proposed method is unspuervised, which can conduct automated detction operation without any prior information, making it quite suitable for those occasions that have little training samples.(3) Aiming at the drawbacks of the unsupervised detection algorithm in accuracy and real time implement, a semi-supervised detection algorithm based on dictionary learning is proposed. This method is in a sense reminiscent of the classical template matching methods but using adaptive templates modeled by the learned dictionary, not needing to consider the alignment between the templates and sample images. SVDD is also introducted to classify detects. Since the propsed method makes detection operation in a more natural way of template matching, the performance of the proposed method does not depend on the type of defects, but the anomaly a defect can cause.(4) Aiming at the real time implement issue limited for most current detection algorithms, the effect of the learned dictionries using different constraints is investigated, sugguesting the orthogonal dictionary has the best overall performance detection accuracy and real time implement. A real time fabric defect detection experiment on160m using orthogonal dictionary is conducted, achieving89%of CDR and3.9%of FAR in the speed of30m/min.