This doctoral thesis focuses on developing an intelligent tool for analysing the content of specific audiovisual production by means of modern artificial intelligence techniques. It answers the growing demand for high-quality automated analysis of artistic audiovisual pieces whose distinctive features render conventional analytic procedures inadequate and time-consuming.
The primary objective is to design a robust classification methodology for both visual and audio content and to implement an interactive web-based application that allows art historians to examine bibliographic and content data efficiently.
The study also introduces a thoroughly annotated data set with specially defined categories for training and validating advanced CNN models. The experimental section evaluates the performance of a pre-trained CNN together with VGG, Inception, Xception and NASNet architectures fine-tuned on the curated data set. For audio analysis, CNN 14, ResNet and Wavegram networks are employed. The results verify the suitability of the chosen methods for the data under investigation.
Finally, the work delivers a comprehensive, user-friendly tool that supports interactive exploration of audiovisual material and provides a dedicated interface for detailed data set construction.