Deep learning based automatic defect detection system in textile production

Fabric defect detection is an important area in textile quality assurance that aims to detect irregularities or abnormalities in textile materials to maintain production standards. The detection of these defects started with manual inspections. With the widespread use of computer vision, defect detection algorithms have started to be produced and their performance is increasing day by day. Simultaneously, pattern recognition algorithms supported by statistical methodologies and feature extraction have been able to distinguish between normal and defective fabric patterns. In recent years, the integration of deep learning techniques, in particular Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), has revolutionised fabric defect detection, offering significantly higher levels of accuracy, efficiency and scalability in the analysis of complex patterns and anomalies. Yet challenges remain in this area, including variability inherent to fabric typologies, distortions in ambient lighting and scalability considerations. In this thesis, a new dataset (YVS) was created for basic fabric defect samples and the TILDA dataset was additionally utilised. The images in the YVS and TILDA datasets have a size of 416x416. In order to eliminate the differences between the high-resolution camera and the low resolution camera, to see the defects on the fabric in detail and especially to prevent small-sized defects from being missed, the images were enlarged by a factor of 4 with the enhanced super-resolution generative contention network (ESRGAN) model. In order to perform this process, the images were first reduced by a factor of 4 and new super-resolution images trained with the ESRGAN model were obtained. Using the model weights obtained at the end of the training, new super-resolution images of 1664x1664 size were obtained by enlarging the original images of 416x416 size by 4 times in the test phase. Thus, small errors became more visible. Then, the 1664x1664 image was divided into 16 local images of 416x416 size as a 4x4 matrix. The images divided into local regions were subjected to the error detection process with the YOLOv8 model, respectively, and then the test results were recombined and the errors on the original image were reported on the global image. For YOLOv8, a segmentation method was used to increase the accuracy of the labelling process and to correctly identify the error. Training and testing procedures were performed separately for TILDA and YVS. The bounding mask mAP50 values were 74.4% for TILDA and 92.2% and 70.7% for dark and light fabric in YVS, respectively. In addition, the ESRGAN and YOLOv8 training model weights performed with the TILDA dataset were also tested in the YVS model. In this case, the bounding box and mask delimiter achieved a mAP50 value of 99.5% for the dark coloured fabric dataset, while a mAP50 value of 90% was obtained for the light coloured fabric dataset. It is concluded that the proposed system can successfully detect defects in plain coloured fabrics. In future studies, it is necessary to carry out researches for the detection of defects in fabrics with fabric motifs and complex patterns.