Optimizing waste reduction through real-time defect detection in garment sewing using variants YOLOv7

Abstract

In the era of computer vision to overcome challenges, the introduction of the YOLO model revolutionized real-time computer vision approaches. In the garment industry, the inception of products plays a significant role while increasing the processing time with a good accuracy rate is the big challenge here. A real-time garments defect detection approach using YOLOv7, YOLOv7x, and YOLOv7-w6 on a primary dataset is proposed with a good FPS rate and better accuracy. Maximum traditional garments inception approaches focused on end product defects while this model suggests detecting defects in the sewing phase so that the cost of the rejected end product can be optimized by detecting them before a product goes through all the phases. For this our research focuses on three subclasses of Seam, Stitch, and Hole related to sewing phase defects. To increase the detection rate, the hyperparameter tuning technique is applied to the YOLOv7 model. Three models are proposed based on pre-trained weights of YOLOv7, YOLOv7x, and YOLOv7-w6 to compare the accuracy and FPS rate in terms of implementation in real-world projects.