2023-10-12 –, Bazar
Many images on Europeana.eu are in low resolution and don't meet user expectations or the requirements set by the Europeana Publishing Framework. Images often suffer from degradations including camera blur, sensor noise, sharpening artifacts, and JPEG compression.
Image Super-Resolution (SR) techniques aim to enhance the visual quality and refine the details of low-resolution images by generating corresponding high-resolution versions with finer details. Recent advances in SR rely heavily on deep learning models that leverage data-driven approaches to accurately reconstruct missing details. However, assessing the quality of the SR output is challenging, as commonly used quantitative metrics like PSNR and SSIM only loosely correlate with human perception.
Different deep learning models may perform better on specific types of images, such as photos, drawings, and prints. Following the experiments of the Europeana Foundation R&D team, at EFHA, in order to determine the most suitable SR model for different image types, we developed a framework for evaluating various deep learning models using selected images from the EFHA collections on Europeana.eu. Through a crowdsourcing visual comparison tool, users have the ability to browse the EFHA test collection, zoom in on different areas of the images and compare and rank the results obtained from different SR models.
By leveraging user feedback and incorporating visual comparison tools, our framework aims to improve the quality of images within Europeana.eu and provide insights into selecting optimal SR models based on image type. Join this session to find out more.
Henk Vanstappen is co-owner of Datable, whose mission is to help heritage institutions better manage, enrich, publish and preserve their digital collections. He has a master in History and Information Sciences and was a technical partner in a.o. Europeana Fashion, Crafted and AI4Culture.
