The project aims to make restoration models more transparent and interactive, and then applying them to real historical photographs.
First, the project developed an AI model that removes different types of image damage — such as blur, noise, low resolution, and poor lighting — using a modular design where each part of the model handles one specific type of damage. This design makes it possible to understand which part of the model is responsible for each improvement, achieving a restoration quality better than existing state-of-the-art methods. This work was published in a scientific journal (Computer Vision and Image Understanding).
The project then developed a similar approach for image colourisation, where the model can explain its colour choices both by linking them to what the user asked for and by relating them to the content of the image itself. A study with real users confirmed that these explanations were understandable in the large majority of cases.
Building on this, the project designed a way for users to actively guide the colourisation process, rather than only receiving an automatic result. Users can draw simple strokes on the image to indicate a desired colour and area, or type a short text instruction (e.g. "red car"), and the model incorporates this guidance directly into the colourisation process. A user study confirmed that this interactive approach worked as intended.
Finally, all these developments were combined into a single, integrated software tool for restoring and colourising historical photographs, allowing users to interactively guide the process and see explanations for the results. In parallel, the project proposed a new way to measure the quality of restored images, which was published at one of the top international conferences in computer vision (CVPR), providing the research community with a more reliable tool for evaluating this type of technology.
All planned research objectives and results were achieved during the project.