Automatic detection and classification of fabric defects with deep learning method

Nowadays, the textile industry has become a sector that is constantly growing and its importance increases day by day. The production of quality and defect-free fabrics is of great importance for both manufacturers and consumers. However, it may be inevitable for errors to occur during the fabric production process, which both increases the cost and negatively affects the quality of the products. Therefore, detecting and intervening in fabric defects at an early stage has become a critical need for the textile industry. In this thesis study, it is aimed to develop an effective defect detection system with an artificial intelligence method based on YOLO for a fabric data set consisting of errors such as ball head error, oil stain error, thread error and hole error. In this study, a data set that has not been used before was collected. The data set was created specifically for the detection of fabric defects and another data set published in the existing literature was also used. By combining both data sets, a new data set was obtained from 5590 images. First, labels for the error class are specified for all images in the data set. Then, to improve the performance of the YOLO algorithm, only the images in the training data set were synthetically reproduced using data augmentation methods. A number of experimental studies were carried out with the Python program to evaluate the ability of YOLO algorithms to detect fabric defects. According to the experimental analyses, the mean avrage precision (mAP), precision (P) and recall (R) performance measures of the models were obtained. YOLOv5s, YOLOv5m, YOLOv5l, YOLOv5x, YOLOv7 and YOLOv8s algorithms were used in the experiments. According to the experimental results, the results given by the YOLOv5s, YOLOv5m, YOLOv5l, YOLOv7 and YOLOv8s models are as follows; %84, %84.5, %86, %85, and %83 mean avrage precision (mAP), %87, %87, %87.2, %86, and %83.2 precision (P), %83, %83.2, %84.6, %84 and %83.2 recall (R) values were obtained. As a result of the studies, the Yolov5x model gave the best results, reaching %86 mean average percision, %87.6 precision and %85.1 recall values. In the conclusion part of the study, the result of YOLOv5x was compared with the results of other models and the improvement values were calculated. In the future parts of this study, YOLO runs and transactions are mentioned in detail.