Image-based anomaly detection in manufacturing: An embedded system application supported by artificial intelligence using textile data

This thesis aims to develop a solution to the problem of image-based anomaly detection for the automatic detection of surface defects encountered in textile production lines using deep learning-based methods. In this context, 9 different models and backbone architectures were systematically evaluated on the MVTecAD and WFDD data sets widely used in the literature; comprehensive comparisons were made using multi-dimensional performance criteria such as accuracy, inference time, memory consumption and number of parameters. In line with the findings obtained, the three most successful models were determined, and the GLASS architecture was preferred as the final choice. One of the original contributions of the study is the creation of a special data set called AIORCOM-TextileAD, which reflects industrial production conditions. The model, which was retrained with this data set, was first tested on high-performance systems and then transferred to the embedded system environment (Raspberry Pi 5). In the tests conducted on the embedded system, the real-time operability of the model and its performance on limited resources were analyzed. The results obtained demonstrate that the selected model can operate with significant accuracy and processing time even on low-cost hardware, offering significant application potential for the integration of artificial intelligence-based automatic quality control systems into the textile industry.