R. Augustauskas “Detection of textured surface defects using deep neural networks” doctoral dissertation defence

Author, Institution: Rytis Augustauskas, Kaunas University of Technology

Science area, field of science: Technological Sciences, Informatics Engineering, T007

Scientific Supervisor: Assoc. Prof. Dr. Arūnas Lipnickas (Kaunas University of Technology, Technological Sciences, Informatics Engineering, T007)

Dissertation Defence Board of Informatics Engineering Science Field:
Prof. Dr. Rytis Maskeliūnas (Kaunas University of Technology, Technological Sciences, Informatics Engineering, T007) – chairperson
Prof. Dr. Nikolaj Goranin (Vilnius Gediminas Technical University, Technological Sciences, Informatics Engineering, T007)
Prof. Dr. Olga Kurasova (Vilnius University, Natural Sciences, Informatics, N009)
Prof. Dr. Renaldas Urniežius (Kaunas University of Technology, Technological Sciences, Informatics Engineering, T007)
Assoc. Prof. Dr. Marcin Wozniak (Silesian University of Technology, Poland, Technological Sciences, Informatics Engineering, T007)

The dissertation defence was at the Meeting room at Santaka Valley of Kaunas University of Technology (K. Baršausko 59 – A228, Kaunas).

The doctoral dissertation is available at the library of Kaunas University of Technology (K. Donelaičio g. 20, Kaunas).

Annotation:

The inspection of manufacturing processes has become an essential part of industry 4.0 nowadays. Quality assessment at each step of production enables the detection of flaws at early fabrication stages, reducing materials usage, thus cutting manufacturing costs. Furthermore, it mitigates the risk of defects appearing in sold production. Non-invasive check-ups, such as those computer vision (CV) based, might be suitable for most observable defects, as the visual-based approach is routinely employed for defect detection in the industry. In this work, deep learning-based approaches are discussed for complicated surface abnormality of visual or structural consistency analysis. Lightweight convolutional neural network designs and model training approaches are proposed to increase the prediction performance while considering computational performance. The developed methods are investigated in artificially generated surfaces, concrete and asphalt defects, and wooden furniture board drilling segmentation datasets.