Fabric Detection Using YOLOv5-Transformer Models

Abstract

Detecting fabric defects, especially in textiles with complex textures, presents significant challenges due to the intricate nature of fabric patterns. Among various object detection methods, the YOLO algorithm is renowned for its real-time performance and accuracy. By treating object detection as a single regression problem, YOLO predicts bounding boxes and class probabilities from an entire image in one pass. This paper proposes an enhanced YOLOv5-Transformer model that integrates Transformer layers into the YOLOv5 architecture to improve fabric defect detection. Performance is evaluated using precision, recall, mAP, and FPS metrics under CPU/GPU environments. YOLOv5, a fully convolutional neural network, strikes an optimal balance between speed and accuracy in end-to-end detection tasks. Leveraging the latest advancements in deep learning, YOLO models deliver high-speed object detection while maintaining competitive precision, making them suitable for real-world applications. Our proposed YOLOv5-Transformer model surpasses other YOLOv5 variants, achieving an accuracy of 82.9%, representing a 5.6% improvement over YOLOv5s and YOLOv5n and a 2.7–3.2% improvement over different versions of YOLOv5m, YOLOv5l, and YOLOv5x. Comparative performance metrics are also presented, including processing time on CPU versus GPU, precision, recall, and F1 score.