|Title:||Development of a computer vision model for quality inspection in textile industry|
|Advisors:||Wong, Calvin (ITC)|
Lai, Zhihui (TC)
|Subject:||Computer vision -- Industrial applications|
Textile industry -- Data processing
Textile industry -- Quality control
Hong Kong Polytechnic University -- Dissertations
|Department:||Institute of Textiles and Clothing|
|Pages:||xviii, 146 pages : color illustrations|
|Abstract:||Advanced development of artificial intelligence (AI), together with the current data-rich environment, has facilitated academic research on computer vision (CV) in the textile and apparel industry. Despite related CV works aimed to address various problems in textile manufacturing, studies that have focused on developing a complete CV model for defect detection and classification on various fabric types are limited.|
The literature on fabric defect detection and classification expresses three concerns. First, most of the existing methods are developed to solve the defect detection problem on specific fabric types. These methods usually present better performance on solid colour fabrics, but the performance on other fabric types cannot be guaranteed. Second, as most studies aim to solve the defect detection problem, attention paid to defect classification is insufficient. Defect classification following the process of defect detection is also an important and valuable research topic because it can help to track the cause of defect generation for quality inspection in the textile and apparel industry. Although some methods have been proposed for classification, they face difficulty in guaranteeing the performance on complex fabrics such as multicoloured or patterned fabrics. Third, the lack of large fabric defect datasets in the literature limits the development of CV algorithms to solve practical problems in textile manufacturing.
In this thesis, an intelligent CV model is developed to deal with the problems of defect detection and category identification in the textile and apparel industry. The model provides corresponding solutions to address the aforementioned concerns. First, the model can obtain high defect detection precision on woven fabrics in solid colours. Furthermore, the model can improve the detection performance on fabrics with regular patterns, such as plaid and dotted patterns. The model uses a weighted double low-rank decomposition (WDLRD) method to locate the defective regions and reconstruct the background of the fabrics. The noise interference on the fabrics is also reduced by the guidance of a prior map in the optimisation procedure. The optimal solution is obtained by employing an alternating direction method of multipliers (ADMM) algorithm. WDLRD is verified to be effective on detecting defects on solid colour fabrics as well as regular patterned fabrics. Second, the model not only releases the problem of defect classification on fabrics with various patterns and colours but also provides new insights to enrich the research on fabric defect classification in the literature. Specifically, a method called jointly sparse locality regression analysis (JSLR) based on sparse representation and locality preserving learning is developed and evaluated on general image datasets. Experimental results show that JSLR can obtain higher classification accuracy than several classical methods. However, JSLR and other general methods encounter difficulty in obtaining satisfactory classification performance on fabric defect datasets. To solve this problem, the author incorporates JSLR with deep convolutional neural network (CNN) and further proposes a new method called Deep-JSLR to learn high-level representations of the data and pursue better classification performance. Deep-JSLR first uses the CNN as a feature extractor to learn high-level representations and then, the performance of those representations are evaluated by JSLR with quantitive loss. Deep-JSLR can obtain higher classification accuracy on fabrics with various types and patterns than several related traditional and deep methods. Additionally, a hashing method called concentrated hashing learning with neighbourhood embedding (CHNE) is designed for defect classification. CHNE considers the similarity of data pairs and simultaneously controls the classification loss during the training process to learn optimal binary representations of the features. Compared with continuous representations of the features, the learned binary representations are more compact for storage and more time-efficient for calculation of defect classification in the textile manufacturing industry. Third, two new fabric defect datasets collected from online open competitions of industrial intelligence innovations are constructed and used in training and evaluating the classification or detection performance of different algorithms.
Overall, the research in this thesis has enhanced the understanding of fabric defect images and provided new insights to widen the scope of research methodology for fabric defect classification. The proposed model can deal with the defect detection problem on solid colour and regular patterned fabrics. The problem of defect classification is also released and the efficiency is improved by incorporating the deep hashing technique. Two new datasets have been constructed for training and evaluating the computer vision algorithms for fabric defect detection and classification. Owing to the complexity of fabric images, we need to solve two main problems in the future. One is the construction of large-scale defect datasets with various fabric types. The other is the development of more powerful models for handling detection and classification tasks on fabrics with complex patterns and various colours.
|Rights:||All rights reserved|
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item: