Metodologia e plataforma baseadas em aprendizado de máquina para inspeção de defeitos na indústria têxtil

Abstract

Quality inspection (QI) in the textile industry is an essential yet challenging process, particularly due to the diversity of defects and reliance on subjective criteria from human inspectors. This work proposes a methodology based on computer vision and deep learning to optimize the QI process in textiles, alongside the development of an associated platform to demonstrate its feasibility and applicability in industrial environments. The platform integrates the proposed algorithms: grid-based detection, which segments images into patches for local defect classification, and ripple refinement, which improves detection quality by analyzing neighboring cells. The experimental results validate the effectiveness of the methodology. A benchmark of machine learning models was conducted, comparing eight approaches, including YOLOv5, YOLOv8 variants, and popular networks in the literature, using the TILDA 400 dataset, composed of five distinct defect classes. The YOLOv8 models stood out, especially YOLOv8 medium, which achieved 90.35% accuracy with an inference time of 0.5ms on a Tesla P100 GPU, surpassing the average precision of 70% from human inspectors. The YOLOv8 small also demonstrated remarkable performance, with a theoretical inspection capacity of up to 46.875 m/min and 86.96% accuracy, exceeding the manual maximum speed of 15 to 20 m/min. The application of this methodology, combined with the developed platform, demonstrates potential to enhance accuracy, speed, and organization in QI, while also reducing operational costs and promoting automation and efficiency in the textile industry.