​Development of a deep neural network for vision-based quality control in nonwoven-industry packaging machines

This thesis presents the development of a vision-based quality control system for monitoring the sealing and welding process in automatic packaging machines. The objective is to detect defects in real-time by leveraging deep learning techniques, primarily based on ResNet architectures. The study encompasses the entire pipeline, starting from data acquisition, which involved collecting and preprocessing videos and images to construct a comprehensive dataset representative of real production conditions. Various neural network architectures were explored, comparing different depths and configurations to assess their ability to identify defective sealing accurately, based on the available data. The models were trained and tested using the obtained dataset, and their performance was evaluated based on key metrics such as accuracy, precision, recall, and inference speed. In addition to pure ResNets, alternative solutions were investigated to determine the most suitable approach. The results highlight the potential effectiveness of deep learning in improving quality control, significantly reducing the need for manual inspections while enhancing production efficiency and reliability.