Deep learning application for textile defect detection

The faulty production of fabric negatively affects the economy and prestige of textile companies. Many methods have been developed in the literature for the detection of defects in the fabric. However, there are not enough data sets on faulty and faultless fabrics in the literature. For this reason, in our thesis, it is aimed to investigate the efficiency of new datasets with a group of deep learning models in order to achieve successful classification with high rates. In the study, two datasets, FBRIC-1-mini and FBRIC-2, were created and two different studies were carried out on these datasets. In the first study, EfficientNetv2m, Seresnet152d and Ecanfnetl0 transfer learning models were compared over the FBRIC-1-mini dataset, which has two classes, erroneous and error-free, and focused on increasing the accuracy rate by using ensemble learning. As a result of the experiment, 95.58% accuracy, 99.17% precision, 91.92% recall and 95.41% F1-score were obtained with the ensemble learning method. In the second study, the focus was on the types of defects in the fabric and the FBRIC-2 data set was created. EfficientNet family, ResNet family and our lightweight convolutional neural network model were used to perform multi-classification on the FBRIC-2 dataset and their success rates were compared. At the end of the study, the EfficientNetB4 model from the EfficientNet family performed the classification process with high accuracy, and with this model, 99.62% accuracy, 99.6% precision, 99.6% recall and 99.6% F1 score values were obtained on the test data.