Skip to Content
× You are about to create a new metadata only record. This record does not auto assign a DOI. To allocate a new DOI use the 'Upload data and allocate DOI' option.

MDTD-Art Dataset - Mask Describable Textures Dataset for Artworks

MDTD-Art is a synthetic artwork-degradation dataset designed to model semantically complex damage through semi-transparent texture overlays. It applies 47 texture classes derived from the Describable Textures Dataset to clean artwork images using controllable alpha masks, allowing degradation type and severity to be varied independently. The resulting samples approximate irregular damage patterns such as cracks, stains, fading and material loss across low, medium and high opacity levels. MDTD-Art provides a scalable and controlled resource for studying restoration under partially retained image information and texture-corruption conditions not covered by conventional degradation benchmarks.

Descriptions

Resource type
Collection
Total Items
0
Size
0 Bytes
Contributors
Editor: Shum, Hubert P. H. 1
1 Durham University, UK
Research methods
MDTD-Art is used to benchmark image restoration and image editing models under controlled texture-overlay degradations. Clean artwork images are paired with semi-transparent texture masks at low, medium and high opacity levels, producing degraded inputs with corresponding clean ground truths. This enables systematic comparison of model performance across degradation severities using full-reference image quality metrics, including L1, PSNR, SSIM and LPIPS, while also supporting analysis of prompt sensitivity, reconstruction fidelity and failure modes such as hallucination and identity drift.
Subject
Computer vision
Artificial intelligence
Image processing
Identifier
ark:/32150/r2f7623c66b
Publisher
Durham University
Collection icon

Actions

Items in this Collection

List of items in this collection
  Title Date Uploaded Visibility Action