Real-Time Fabric Defect Detection Utilizing Deep Learning-Based Convolutional Neural Networks

Abstract

Considering the reduction in labor costs and associated benefits, investment in automated fabric defect detection proves to be highly cost-effective. A robust and efficient algorithm is essential for the development of a fully automated web inspection system. The examination of genuine fabric flaws is particularly difficult because of the multitude of defect categories, which are defined by their ambiguity and indistinctness. This work aims to classify and describe several strategies developed for detecting fabric faults. This work also provides the inaugural survey on methodologies for fabric defect detection, referencing around 160 sources. The categorization of fabric defect detection methodologies is beneficial for assessing the characteristics of the identified features. The characterization of authentic fabric surfaces through their structure and primitive set has not yet demonstrated success. Consequently, the characteristics derived from fabric surfaces have led to the classification of the proposed methodologies into three categories: statistical, spectral, and model-based. To evaluate the state-of-the-art, the constraints of several prospective techniques have been identified, and their performance has been appraised based on demonstrated findings and proposed applications. The results of this work indicate that integrating certain statistical, spectral, and model-based methodologies may produce superior outcomes compared to any individual strategy, warranting additional investigation into this matter.

INTRODUCTION

Fabric quality inspection plays a crucial role in the textile and yarn industries. Surface imperfections can significantly impact garment quality, leading to customer dissatisfaction and potential losses [1]. Therefore, there is an urgent need for a high-performance, fast, and reliable automated inspection system [2]. Currently, most textile manufacturers rely on manual defect identification, which

is time-consuming and inefficient. Human inspection typically detects only 60-70% of defects in real-time systems [2, 3]. Distractions encompass visual distractions (diverting one's gaze), cognitive distractions (being preoccupied with thinking), auditory distractions (being interrupted by a colleague), and weariness during garment testing. Consequently, human agility is crucial as distractions can result in considerable losses [4, 5]. Furthermore, obstacles have escalated from an industrial perspective [6]. The industry must uphold both the quality and quantity of production to preserve its credibility in the market. Consequently, people are unable to do this inspection work for extended periods while maintaining the necessary precision.

Textile manufacturing companies globally require automation of the inspection process [7]. This inspires us to address this complex issue as a research statement. Digital image processing is utilized for the processing, storage, and display of human-interpretable visual information.

Consequently, computer vision, along with MATLAB, OpenCV, and Keras, is essential and demonstrates efficacy in addressing issues posed by the contemporary industrial landscape [8]. A portion of the fabric that fails to satisfy the specified standards or characteristics is classified as a defect, leading to customer unhappiness and potentially significant losses for the textile industry [9]. Various sorts of fabric flaws include missing yarn, broken yarn, double yarn, stains, and holes. Occasionally, the yarn becomes entangled during preparation, leading to holes, which is the most unattractive flaw. The machine's malfunction may result in problems, including absent yarn. A significant disadvantage is the presence of stain marks [10].

Deep learning is a cutting-edge approach utilized for the detection of flaws in fabric. The system employs a high-resolution camera for picture capture and a stepper motor for the conveyor mechanism. Moreover, enough illumination is crucial, as insufficient lighting can result in substandard digital photos. The suggested system will employ a neural network to recognize and categorize the photos. The implementation of neural networks can enhance system performance. Processing speed can be improved using a high-performance GPU or CPU, hence decreasing computing time [11].

Deep learning is an advanced approach utilized for identifying faults in cloth. The system employs a high-resolution camera to capture images and a stepper motor for the conveyor mechanism. Furthermore, an adequate illumination source is essential, as insufficient lighting may lead to substandard digital photos. The suggested system would employ a neural network to recognize and categorize

photos. Neural networks enhance system efficiency. Utilizing advanced GPUs or CPUs can enhance processing performance and diminish computing time [12].

RELATED WORK

An efficiently automated system facilitates reduced labor costs and minimizes manufacturing time. This issue draws considerable attention from researchers, resulting in numerous articles that have developed fabric flaw-detecting systems employing diverse methodologies. Ouyang et al. [13] have investigated the detection of fabric defects through the utilization of an embedded convolutional neural network activation layer, employing fabric image autocorrelation to define fabric motifs and leveraging it as a prominent feature. They forecasted elevated precision and introduced PPAL-CNN.

Hanbay et al. [14] introduced an optical-based fabric fault detection method. The quantity of yarns and fibers functions as essential units. Nonetheless, prospective fabric imperfections can be mitigated by evaluating yarns and fibers prior to fabric production. Wenninger et al. [15] suggested a flaw detection system for plain woven fabrics utilizing a fully convolutional network alongside yarn tracking. A method for fabric fault identification has been described. The fibers of the fabric might be identified and monitored without any conditional parameters. Subsequently, the ripple faults were identified. This issue was previously addressed using a neural network, but was limited to grey fabric exclusively. This research introduced a sophisticated colored fabric defect detection system utilizing deep learning, featuring a camera positioned above the moving fabric at an appropriate distance, accompanied by an illumination source. The system would collect the image and input it for processing (testing). In the event of an error, the motor will cease operation instantaneously. Consequently, this system would be highly beneficial in both small-scale and large-scale textile enterprises that aim to minimize labor costs while achieving superior fabric quality in the quickest possible timeframe.

Liu et al. [16] conducted a systematic analysis of contemporary variants of low back pain and presented a taxonomy to more distinctly categorize the notable alternatives. The advantages and disadvantages of the different LBP traits and their interrelations were also examined. A comprehensive performance evaluation was conducted for texture classification, empirically analyzing forty texture features, which comprised thirty-two recent and promising LBP variations and eight non-LBP descriptors derived from deep convolutional networks across thirteen widely-utilized texture datasets. The tests aimed to assess their resilience to various classification problems, encompassing alterations in rotation, scale, illumination, viewpoint, class quantity, diverse forms of image degradation, and computational complexity. Hoang and Rebhi [17] examined classical color spaces to assess the impact of LBP accuracy on fabric flaw detection. They achieved 92.1% accuracy in the LUV color space using a classification of 10 classes.

METHODOLOGY

The suggested transfer learning approach employs the AlexNet architecture with suitably initialized weights. The features are subsequently input into the classifier, which predicts the corresponding labels for the vessels. Due to the inferior accuracy of existing approaches, such as Inception and ResNet, compared to AlexNet,...