A Deep Learning based Approach for Automated Defect Detection of Fabrics in Garment Production Line

Abstract

It is estimated that over 50% of garment product rejections are due to raw material defects, which leads to huge amounts of financial loss every year. The rework rate of these industries exceeds approximately 7% on a daily basis, in turn lowering efficiency and productivity. The existing solution to this problem is manual checking for each and every apparel item to either approve or reject the items which is tedious and often ends up with human error. This paper proposes a deep learning based solution to this problem by automating the task of quality control in the aforementioned apparel industry. By distinguishing between defective and non-defective garment products and localizing the specific type of defect within the fabric, the proposed model makes the task much more efficient and streamlined and therefore, helps to redirect the defected products to its specific market, even in the worst case complete rejection. The approach is based on the classic object detection model, YOLOv8, along with necessary extensive image preprocessing measures and hyperparameter tunings. For training and validation purposes, a custom dataset consisting of over 1500 garment product images of multiple classes has been prepared. Various classes of defects, including horizontal or vertical lines from faulty stitching, tears, or spots within the fabric, were collected and superimposed on top of the gathered images, and the final image was rendered to be included within the dataset. The use case is further strengthened as the model has managed to make accurate detections when tested with real world non-synthetic images. It has proven to be competitive in the field of garment defect detection, managing to achieve a staggering 97.96% mean average precision.