This doctoral dissertation focuses on the development of a deep learning system for visual quality inspection of products in mass production. The main goal of the research was to develop a model for anomaly detection that enables incremental learning, thereby reducing the need for time-consuming expert image annotation.
The dissertation presents the development of the FuseDecode AE model, which, based on the reconstruction of normal instances, enables anomaly detection through various learning phases, from unsupervised to semi-supervised and mixed-supervision learning. With each subsequent phase of learning, the model achieves more reliable anomaly detection and segmentation. One of the key advantages of the proposed model is the reduction in the time required for annotating training datasets for semi-supervised and mixed-supervision learning, as the unsupervised learned model serves as a tool to accelerate the process. The methods were validated on both a real industrial dataset of pipes coated with KTL coating and the publicly available MVTec AD dataset, where FuseDecode AE achieves results comparable to other state-of-the-art anomaly detection methods.
|
