Deep learning based texture defect detection in manufacturing systems

Recently, concepts such as internet of things, object classification, object detection, and pattern recognition have become quite common today with the developments in computer software and hardware technology. Systems developed with these concepts have provided innovation and convenience in many fields such as medicine, chemistry, engineering, agriculture, security systems, and geographical sciences. Especially deep learning-based systems have provided high performances in many areas. This thesis focuses on adapting existing deep learning systems to manufacturing systems and developing new approaches for automatic texture defect detection in manufacturing systems. Generally, surface/texture defect detection in manufacturing systems is performed manually by expert personnel. Manual inspections have downsides in terms of time and accuracy. In addition, it is costly due to the need for expert personnel. In comparison, automatic defect detection systems supported by artificial intelligence can produce more successful results in terms of performance and reliability. Thanks to these successful results, financial losses caused by manual management can be prevented and strong profits can be achieved in the field of production. In this context, new and effective models based on deep learning have been proposed for automatic texture defect detection in this thesis. These models are listed below: • Feature-enriched pyramid network architecture for defect detection, • Depth-wise Squeeze and Excitation Block (DSEB) to strengthen important features and DSEB-based Efficient-Unet architecture, • A new approach to extracting multidimensional spatial, spectral, and semantic features in deep learning, • Hybrid Attention Gate which developed to adapt the recently very popular Vision Transformer to surface defect detection. • A new encoder and decoder network architecture based on Swin Transformer, • A new approach based on Deep Neural networks, Multiple Pooling, and Filtering for the classification of fabric defects Existing and current datasets in the literature were used to analyze the success of these proposed original methods. In experimental studies, the proposed methods were compared with current models. According to the results of extensive experimental studies, it has been observed that the methods developed for defect detection are effective.