Fabric defects identification for textile industry with a deep learning approach

The textile and clothing industries are crucial sectors that make significant contributions to the economic growth and advancement of countries. In addition to low labor costs, the sector's development performance is influenced by flexible production structure and defect-free products manufacturing. However, in the textile companies, the manufacture, processing, and weaving of fabric can lead to undesirable defects, which ultimately leads to a decline in both quality and cost. This study investigates the effectiveness of different deep learning frameworks to accurately classify fabric defects commonly encountered in the textile industry in Turkey. For this purpose, an innovative data set is generated as comprising fabric defects such as lines, wrinkle marks, machine oil leaks, holes, and bleaching. Furthermore, the effectiveness of Adam and Ranger optimization functions in defect identification has been evaluated using various models in connection with explainable artificial intelligence. The findings indicate that the ResNet18 + Adam model, which is very simple and shallow, obtained a notable level of success with an accuracy of 99.30%. On the other hand, the more complex EfficientNetv2m + Adam model emerged by achieving an accuracy of 99.42%. The research results suggest that deep learning algorithms can effectively replace human manual control in the task of detecting fabric defects.