Classification of Fabric Defects Using Deep Learning Algorithms

The enterprises in the textile sector, which adopt the philosophy of continuous improvement in quality, attach importance to detecting the errors that may occur due to the nature of production. Failure to correctly identify the faults that occur at the very beginning of the production line, that is, during the weaving phase, can cause irreparable material damages and losses in the textile market for businesses in the next stages. The most common method used to detect the defects of woven fabrics is the light board. Quality control personnel control the faults on the woven fabric on this board. However, at this stage, since the detection of errors is made with the human eye, the focus, mood and width of the field of view of the operator who performs the quality control process cause errors to be overlooked. Many methods are used in the detection of fabric defects. These are; statistical, structural, deep learning, spectral, hybrid and comparison methods. In this study, it was decided to use the deep learning method for the detection and classification of the defects of the finished fabrics. The data set used in the study consists of fabric defects obtained from the real production process in a one-month period. On this dataset, it was decided to use Faster R-CNN, SSD, Resnet50, Resnet101 algorithms from deep learning models in the literature. Based on the training times and total loss graphs in the algorithms used, it was concluded that the best method to be used for such a problem is Faster R-CNN.