Significance Automatic fabric defect detection is one of the key aspects of digital quality control in the textile industry. At present, the domestic fabric defect detection is mostly based on manual detection, but the traditional manual detection success rate of only 60%-75%, indicating that the method can't meet the demand for high-quality products. To overcome the drawbacks of manual defect detection, researchers have proposed a variety of learning-based defect detection algorithms. Compared with the manual detection, machine learning methods demonstrate a high detection rate, good stability and other characteristics. Bacause of the superiority of deep learning technology in defect detection, this technology is also used for fabric defect detection. In order to improve the efficiency of the application of deep learning technology in defect detection and to achieve digital quality control in the textile industry, the current Status of researeh on deep learning technology in defect detection is discussed.Progress Although traditional algorithms have achieved imroved results in some specific applications, there are still limitations when dealing with complex fabric textures. With the upgrading of Computer hardware, the technology is superior in the fields of target detection and image Classification, and is utilized in textile quality inspection. Since the introduetion of deep learning, great breakthroughs have been made in target detection, which can be categorized into one-phase detection model and two-phase detection model in textile defect detection, both achieving better results in detection speed and detection aecuraey. Due to the excellent feature extraction capability of neural networks, convolutional neural network (CNN) based Classification networks are widely used for surface defect detection and Classification, which can automatically learn different types of fabric defects and accurately categorize them into different classes. Various deep learning methods are superior to manual detection. Due to the difficulty in obtaining fabric datasets, researeh based on unsupervised learning and semi-supervised learning is gaining popularity, which trains on unlabeled data and a small amount of labeled data and reduces the dependence on labeled data. It can effectively deal with unlabeled datasets or situations where labeled data is scarce or unavailable, and it greatly reduces the working time compared to supervised learning where training is performed on labeled datasets.Conclusion and Prospect This paper reviews the application of deep learning techniques to fabric defect detection. First, publicly available defect datasets are organized and analyzed. Secondly, the principles, advantages and disadvantages, and the scope of application of deep learning techniques for defect detection are summarized from three perspectives, i. e. supervised learning, semi-supervised learning and unsupervised learning. In addition, the commonly used speed and aecuraey evaluation metrics in defect detection are sorted out. Finally, the experimental results of different deep learning networks in the detection task are objectively compared and analyzed, and the future development direction of fabric defect detection is envisioned. Supervised learning-based defect detection requires a large number of datasets, and the available public data resources are relatively scarce. Relying solely on manual labeling of fabric defects is not only time-consuming but also inefficient, therefore, automatic labeling of fabric defects and detection methods that do not require data labeling have become an important direction for future researeh. Currently, defect samples face many challenges in terms of data scarcity, labeling difficulty, and uneven data distribution, so unsupervised learning, weakly-supervised learning, zero-sample learning, and small-sample learning are reeeiving more and more attention in defect generation and detection. On the other hand, solving the data problem and developing defects with fabric texture characteristics is also one of the focuses of future researeh. Currently, most network struetures are still designed manually. However, with the development of automatic machine learning techniques, more and more machines will be able to search and generate network architectures automatically, gradually replacing the traditional manual design. © 2024 China Textile Engineering Society. All rights reserved.
