Machine learning for analogue media damage restoration

Ivanova, Daniela (2025) Machine learning for analogue media damage restoration. PhD thesis, University of Glasgow.

Abstract

Analogue media degradation presents a unique challenge for digital restoration, requiring the disentanglement of physical damage from intended content and style. This thesis investigates machine learning approaches for detecting and segmenting damage across diverse media types, addressing a critical bottleneck in restoration workflows.

We first focus on the constrained setting of a single type of analogue media, and develop a statistical framework for modeling film damage. Through perceptual studies, we demonstrate that our approach generates training data indistinguishable from authentic damage. Using this data to train supervised models yields significant improvements in damage detection performance, as validated on our benchmark dataset of professionally restored high-resolution film scans.

Expanding beyond film, we introduce ARTeFACT, the first comprehensive dataset for analogue media damage detection comprising over 11,000 pixel-accurate annotations across 15 damage categories and 10 diverse media types. Systematic evaluation reveals that state-of the-art supervised segmentation methods, including foundation models like Segment Anything, fail to generalize across different media and struggle to disambiguate damage from visually similar content features.

To address these limitations, we investigate the semantic understanding capabilities of text to-image diffusion models. We develop a novel unsupervised , zero-shot semantic segmentation framework leveraging self-attention mechanisms in Stable Diffusion, achieving state-of-the-art performance on standard benchmarks. Through matrix exponentiation of attention maps, we provide a principled control mechanism for segmentation granularity based on random walk theory. Such mechanism is crucial for enabling segmentation of artefacts at varying granularity.

Our findings demonstrate that effective analogue media damage detection requires moving beyond rudimentary pattern recognition toward semantic understanding of content-damage relationships. This work establishes a methodological foundation for automated restoration systems that can support heritage preservation at scale while respecting the unique characteristics of different analogue media. The challenges presented by damage detection further illuminate fundamental questions about representation learning and visual-semantic disentanglement that are significant for advancing machine learning beyond pattern recognition toward more meaningful visual understanding.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science; Q Science > QA Mathematics > QA76 Computer software
• Colleges/Schools: College of Science and Engineering > School of Computing Science
• Supervisor's Name: Williamson, Dr. John and Henderson, Dr. Paul
• Date of Award: 2025
• Depositing User: Theses Team
• Unique ID: glathesis:2025-85461
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 23 Sep 2025 13:27
• Last Modified: 23 Sep 2025 13:34
• Thesis DOI: 10.5525/gla.thesis.85461
• URI: https://theses.gla.ac.uk/id/eprint/85461

Related URLs

• Article DOI
• Conference item
• Pre-print
• Conference proceeding
• Article DOI
• Enlighten Publications Record
• Pre-print
• Book section
• Pre-print