Date created: | Last Updated:

Category: Communication

Description: Presented at JTS2019 by Franz Hoeller at the Netherlands Institute for Sound and Vision, Hilversum (NL) on Saturday, October 5, 2019. Collaborative notes available at: https://doi.org/10.5281/zenodo.3835666 ABSTRACT: DeepRestore is a research project driven by HS-ART Digital (Austria) together with the TU-Graz Institute of Computer Graphics and Vision. (Austria), running from 01/2018 until end of 2020. The goal of the project is to evaluate AI technologies in the sector of digital film restoration. A focus is on applying modern machine learning techniques, in particular convolutional neural networks, to remove dust and scratches in archival footage. In the presentation we will show current results of the DeepRestore prototype and we will compare the speed and quality of the restoration with the classic dust & scratch filters from the DIAMANT-Film Restoration Software. We will discuss the advantage and disadvantage of the AI approach versus the classical approach in digital film restoration. The problematic of generating good training data for the particular problem to solve will be shown and possible solutions will be presented.