Defect detection in clothing remains a barely addressed topic on the literature. However, if the scope of the topic is expanded to the industry, some interesting works have been carried out, mainly regarding the fabric quality control in the textile industry. Such quality control approach still plays an important role in the industry, and can be an appealing starting point for defect detection in clothing with other purposes in sight [17].

Based on the aforementioned premise, a quick literature survey allows perceiving that machine vision based on image processing technology has replaced manual inspection, and allows for reducing costs and increasing the detection accuracy. An integral part of modern textile manufacturing is the automatic detection of fabric defects [18]. More recently, due to their success in a variety of applications, deep learning methods have been applied to the detection of fabric defects [19]. A wide range of applications were developed using convolutional neural networks (CNNs), such as image classification, object detection, and image segmentation [20]. Defect detection using convolutional neural networks can be applied to several different objects [21,22,23]. Comparatively to traditional image processing methods, CNNs can automatically extract useful features from data without requiring complex feature designs to be handcrafted [24]. Zhang et al. [25] presented a comparative study between different networks of YOLOv2, with proper optimization, in a collected yam-dyed fabric defect dataset, achieving an intersection over union (IoU) of 0.667. Another method, unsupervised, based on multi-scale convolutional denoising autoencoder networks, was presented by Mei et al. [26]. A particularity of this approach is the possibility of being trained with only a small number of defects, without label ground truth or human intervention. A maximum accuracy of 85.2% was reported from four datasets. A deep-fusion fabric defect detection algorithm, i.e., DenseNet121-SSD (Densely Connected Convolutional Networks 121-Single-Shot Multi-Box Detector), was proposed by He et al. [27]. By using a deep-fusion method, the detection is more accurate, and the detection speed becomes more efficient, achieving a mean average precision (mAP) of 78.6%.

Later, Jing et al. [28] proposed a deep learning segmentation model, i.e., Mobile-Unet, for fabric defect segmentation. Here, a benchmark is performed with conventional networks on two fabric image databases, the Yarn-dyed Fabric Images (YFI) and the Fabric Images (FI), allowing to reach IoU values of 0.92 and 0.70 for YFI and FI, respectively. A novel model of a defect detection system using artificial defect data, based on stacked convolutional autoencoders, was then proposed by Han et al. [29]. Their method was evaluated through a comparative study with U-Net with real defect data, and it was concluded that actual defects were detected using only non-defect and artificial data. Additionally, an optimized version of the Levenberg–Marquardt (LM)-based artificial neural network (ANN) was developed by Mohammed et al. for leather surfaces [30]. The latter enables the classification and identification of defects in computer vision-based automated systems with an accuracy of 97.85%, compared with 60–70% obtained through manual inspection. Likewise, Xie et al. [31] proposed a robust fabric defect detection method, based on the improved RefineDet. Three databases were used to evaluate their study. Additionally, a segmentation network with a decision network was proposed by Huang et al. [32], with the reduced number of images needed to achieve accurate segmentation results being a major advantage. Furthermore, a deep learning model to classify fabric defects in seven categories based on CapsNet was proposed by Kahraman et al. [33], achieving an accuracy of 98.71%.

Table 1 summarizes the main results of the aforementioned works, including the datasets used.

The results presented in Table 1 demonstrate a lack of standardization in the evaluation metrics and datasets between studies, leading to difficulties in accurately comparing results. This can be attributed to the diversity of tasks in defect detection, including defect classification, defect location, defect segmentation, and defect semantic segmentation, each requiring distinct metrics for evaluation. Furthermore, the studies are focused on one-stage and two-stage detectors, without a comparative study between them. One-stage detectors, such as You Only Look Once (YOLO) [34] and the Single-Shot Detector (SSD) [35], are known for their speed, but also for their lower accuracy compared to two-stage detectors, such as Faster R-CNN [36] and Mask R-CNN (region-based convolutional neural network) [37]. Two-stage detectors offer improved accuracy, but at the cost of a slower performance.

Despite the similarities between clothing and textiles, a new approach is needed for detecting defects in clothing, especially to assist blind people. For that, different types of images must be analyzed, other than just textiles, resulting in the creation of new datasets. In the textile industry, fabrics usually emerge from the manufacturing process in a roll and undergo stretching, augmenting the detection of defects. Furthermore, the magnification of images to fit the fabrics coming off the roll can greatly amplify any defects present, as depicted in Figure 1.

It becomes clear that a comprehensive dataset that captures the entirety of a garment can provide crucial insights into identifying defects in the piece as a whole, thus, leading to significant advancements in this field. Furthermore, textile fabrics’ datasets may not capture important clothing features, such as wrinkles, patterns, and buttonholes, which can present a significant challenge during the analysis, since defects can be hidden in the wrinkles of the clothes, or simply hidden by the way the garment was folded or stored, as illustrated in Figure 2.