Enhancing Fabric Detection and Classification Using YOLOv5 Models ^†

Abstract

The YOLO series is widely recognized for its efficiency in the real-time detection of objects within images and videos. Accurately identifying and classifying fabric types in the textile industry is vital to ensuring quality, managing supply, and increasing customer satisfaction. We developed a method for fabric type classification and object detection using the YOLOv5 architecture. The model was trained on a diverse dataset containing images of different fabrics, including cotton, hanbok, dyed cotton yarn, and a plain cotton blend. We conducted a dataset preparation process, including data collection, annotation, and training procedures for data augmentation to improve model generalization. The model’s performance was evaluated using precision, recall, and F1-score. The developed model detected and classified fabrics with an accuracy of 81.08%. YOLOv5s allowed a faster performance than other models. The model can be used for automated quality control, inventory tracking, and retail analytics. The deep learning-based object detection method with YOLOv5 addresses challenges related to fabric classification, improving the abilities and productivity of manufacturing and operations.

1. Introduction

Deep learning has improved rapidly recently. Image recognition and video classification are widely used within industry to judge the ripeness of fruits and vegetables [1]. Object detection methods have attracted much attention, as exemplified by the YOLO algorithm. The advent of YOLO has notably enhanced detection speeds, expanding the operation of deep learning into various domains.

In online shopping applications, platforms such as Amazon, eBay, and Taobao have seamlessly integrated YOLO for diverse functionalities [2]. In agriculture, YOLO is extensively applied to environmental monitoring, crop health assessment, and pest detection. In online shopping applications, image recognition is used in clothing design, coordinating accessories, categorizing items, and conducting analyses [3]. With the rapid growth of e-commerce, online shopping has provided consumers with convenience and cheaper products. In the apparel industry, engaging customers depends on visual elements such as color, style, and other features, highlighting the need for accurate image representation [4].

For online stores, showcasing garments’ various color and style variations as well as many try-on pictures and accessory options is paramount. This facilitates a straightforward selection aligned with customers’ preferences and enhances the shopping experience [5]. The applications of AI technologies such as YOLO extend beyond online shopping to clothing classification and detection, underscoring their versatility and potential in the industry [6]. Leveraging YOLO’s real-time processing capabilities and high accuracy in object detection, online retailers can streamline the browsing experience, offer accurate product recommendations, and elevate customer satisfaction. Moreover, this integration empowers online retailers to optimize inventory management to enhance the shopping experience.

2. Related Works

Object detection and image classification have been advanced thanks to deep learning techniques. In particular, the YOLO models have garnered considerable attention due to their ability to execute object detection. Video classification using deep learning techniques is used to simulate clothing for customers [7]. Song et al. [8] proposed evolved YOLOv5-based object detection to address inaccurate detection. The enhanced model achieved a better object detection performance using network pruning and parameter optimization [8]. The eye-in-hand measurement method showed increased precision, recall, mean average precision (mAP) value, and F1 score in contrast with the original YOLOv5 models [8]. Li et al. [9] enhanced object detection and addressed illegal immigration, industrial scenarios, natural disasters, and locating missing individuals or items utilizing photography with an F11 4K PRO drone and the VisDrone dataset.

Aduen et al. [10] enhanced the YOLOv5 object detection of smaller objects, particularly in autonomous racing, proposing a modified YOLO-Z model. The model’s mAP was enhanced by 6.9% for fewer objects at an intersection over union of 50% in object detection and its speed increased by 3 ms compared with the original YOLOv5. Hang et al. [11] proposed YOLOv5 for object detection in satellite images by introducing fusion layers and predicting heads. They used swin transformer prediction heads (SPHs) to reduce calculation intricacy and integrate normalization-based attention modules (NAMs) to improve attention performance. Evaluated on the NWPU-VHR10 and DOTA datasets, the modified model SPH-YOLOv5 outperformed the original YOLOv5 with an increase of 0.071 in mAP. Additionally, most YOLO versions enhance object detection based on machine learning.

3. Methodology

YOLOv5 is an advanced deep-learning technique that consists of an input, backbone, neck, and head. Among its variants, YOLOv5s is notable for its compact size, making it ideal for edge artificial intelligence (AI) and deployment on microcontroller units (MCUs). Deep learning models rely on GPU-powered systems for efficient computations. We used an open-source Jupyter Notebook (7.2.2) for model training and testing. Figure 1 shows the steps used for setup, configuration, and running process applications to create a new YOLO layer. Users need to confirm the graphics processing unit (GPU) and compute the unified device architecture (CUDA) settings before installing YOLO algorithms. Finally, the selected YOLO algorithm was trained and tested by using pre-defined weights.

3.1. YOLO Algorithms

YOLO is a single-stage object detection algorithm renowned for its efficiency and effectiveness. Unlike traditional two-stage methods, YOLO formulates object detection as a regression problem. It predicts bounding boxes and class probabilities from entire images in a single evaluation, offering real-time inference capabilities. This approach enables YOLO to achieve a high accuracy while maintaining an impressive speed, making it well-suited for various applications such as autonomous vehicles, surveillance systems, and robotics. YOLO streamlines the object detection process by simultaneously handling localization and classification and provides a robust and efficient solution for computer vision tasks.

The YOLO series has evolved through Darknet powering versions 1 to 4, while YOLOv5 is built on PyTorch (2.5.1+cu121). YOLOv1, inspired by GoogLeNet, treated object detection as a regression task. YOLOv2 improved accuracy and speed by adopting the Darknet-19 framework, though challenges remained in detecting small objects. To address this, YOLOv3 introduced a feature fusion module and expanded its output to three dimensions, enhancing performance with smaller objects. Based on such advancements, YOLOv4 refined its detection accuracy. YOLOv5, the latest version, allows for more compact models.

The YOLO architecture is structured using the backbone, neck, and head. Figure 2 highlights the main differences among the various versions of YOLO. In YOLOv5, the FOCUS function enhances the processing speed by slicing and rearranging the input image. Spatial pyramid pooling (SPP) reduces image distortion in measurements, color, or culture. It accelerates the generation of candidate boxes while decreasing computational overheads.

3.2. Dataset

Image datasets were collected using a camera, capturing data from the fabric category machine shown in Figure 3. The dataset comprised more than 5000 images and was categorized into five groups: plain cotton fabric, the hanbok fabric nobang, dyed cotton yarn, hanbok fabric, and a plain cotton blend fabric.

The dataset was split into training, validation, and testing sets. The training set was used to train the model, while the testing dataset was used for fine-tuning parameters and selecting the best model.

Table 1. Fabric dataset in this study.

Fabric Dataset	Training	Validation	Testing
Plain Cotton Fabric	756	216	108
Hanbok Fabric	945	270	135
Dyed Cotton Yarn	1029	294	147
Hanbok Fabric, Nobang	1197	342	171
Plain Cotton Blend Fabric	1470	402	201

presents the distribution of samples in each category used in this study.

3.3. Image Labeling

LabelImg is an annotation tool developed in Python (3.12.4) that applies PyQt for its graphical interface annotations. It stores XML files in the PASCAL VOC format. Movere is also used in the YOLO and CreateML formats. The tool determines the object’s area and class in labeling and selects an appropriate layout compatible with YOLO. This information includes the class, X, and Y coordinates and the distance and confines of objects Figure 4.

4. Result and Discussion

We used a Windows 10 system with an AMD Ryzen 9 5900X processor running at 3.70 GHz, 128 GB of RAM, and an NVIDIA GeForce RTX. PyTorch, TensorFlow, and other supporting libraries were used to develop the YOLOv5 with a backbone, neck, and head. In object detection, the spatial locations of objects are represented by encircling boxes. YOLO employs anchor boxes of different sizes and shapes to manage overlapping objects within images. Each anchor box is evaluated to determine which object’s encircling box has the highest overlapping-to-non-overlapping ratio. In training, the expected encircling boxes are more continuous than the ground-truth labels, enhancing the detection performance where bounding boxes are color-coded by class, as shown in

Table 2. Performance obtained for 5 and 100 epochs. Time spent by CPU and GPU.

Model Name	mAP@0.5 (%)	mAP@0.95 (%)	CPU Time (s)	GPU Time (s)
YOLOv5s	60.07	81.08	1805 s	665 s
YOLOv5n	45.07	81.02	3364 s	665 s
YOLOv5m	50.07	81.02	1.1169 s	784 s
YOLOv5l	50.08	81.02	1.8475 s	906 s
YOLOv5x	60.33	81.02	3.8888 s	1200 s

and the results in Figure 5.

5. Conclusions

We used the lightweight YOLOv5 algorithm for clothing recognition in various conditions and weight constraints. YOLOv5s, a single-stage object detection approach, enhances both the detection speed and size of the compact model. The model was developed and tested using the Jupyter Notebook in an open-source integrated development environment. Image samples were taken using the LUMIX GH6 camera. Steps for building an efficient and cost-effective computing setup were developed. The experimental results highlighted the benefits of YOLOv5s regarding mAP, CPU and GPU operation times, and efficiency.