Sunday 23 February 2025
Damage detection in analogue media is a crucial step in preserving cultural heritage, but it’s a complex task that has long eluded machine learning models. Researchers have struggled to develop reliable methods for identifying damage in diverse types of analogue media, including paintings, photographs, textiles, mosaics, and frescoes.
A new study aims to address this challenge by introducing ARTeFACT, a comprehensive dataset containing over 11,000 annotations of damage across various subjects, media, and historical provenance. The team behind the project has also developed human-verified text prompts describing the semantic contents of the images, as well as additional textual descriptions of the annotated damage.
To test their approach, the researchers evaluated several deep learning models, including convolutional neural networks (CNNs), transformers, and diffusion-based segmentation models. They found that none of these models performed acceptably in detecting damage, even after being trained on large datasets.
One of the main challenges is that each type of analogue media has its own unique characteristics and requires specialized expertise to identify damage. For example, a painting may require knowledge of art history and techniques, while a photograph might need understanding of photographic processes.
The researchers used their dataset to train models that could learn from diverse types of analogue media and identify damage in various contexts. They found that even the most advanced models struggled to generalize across different materials and contents.
Another challenge is that many analogue media are degraded or damaged over time, making it difficult for machines to accurately detect damage. The team used their dataset to develop techniques for simulating analogue film damage, which can help improve the accuracy of machine learning models in detecting damage.
The study highlights the complexity of damage detection in analogue media and underscores the need for more sophisticated approaches that can account for the unique characteristics of each type of media. While machine learning models have made significant progress in recent years, they still struggle to accurately detect damage in diverse types of analogue media.
Ultimately, developing reliable methods for detecting damage in analogue media will require a combination of advances in machine learning, computer vision, and domain-specific knowledge. By leveraging the strengths of each field, researchers can create more accurate and effective tools for preserving our cultural heritage.
Cite this article: “Challenges in Machine Learning-Based Damage Detection in Analogue Media”, The Science Archive, 2025.
Machine Learning, Analogue Media, Damage Detection, Cultural Heritage, Computer Vision, Domain-Specific Knowledge, Convolutional Neural Networks, Transformers, Diffusion-Based Segmentation Models, Dataset
