Deep Restore
Thanks to the Austrian research structure of FFG , our current research project DeepRestore is the first step into the domain of Artificial Intelligence and Neural Networks for our well known DIAMANT-Film Restoration environment.
The main goal is increasing precision and restoration efficiency, based on Deep Learning models for dust/dirt and permanent defects such as scratches and camera hairs.
DeepDust
In the first year of the project, we worked together with our scientific partner TU-Graz on the rapid prototyping of various neuronal networks, based on Tensor Flow techniques.
The development of a dedicated and specific software supported the manual creation of some 500 verified HQ ground-truth samples for dust &dirt defects and their optimal repair. Supported by mathematics and generalization, the ground-truth forms the basis of a representative collection of 18,000 training samples.
Training Data
Results
After a full training with the full set of training data, the following exemplaric results with the AI based prototype filter could be obtained:
The above image shows the result of the DeepDust in a typical detail. The repair results look very promising. The natural grain structure is fully preserved by the selected approach. The main limitation is the detection, where a significant number of false positives indicates the difficulty to distinguish between moving objects (e.g.: leaves and grass going with the wind) and dirt in front of it.
Watch a full before/after video here.
Inpainting
We have looked into severs inpainting methods using AI technologies. See some results here.
In the sample Big Scratch which was an 4K scan with a lot of grain in the picture we have to say that the AI inpainting result is not very good. The result is bury. The conventional algorithm in DIAMANT-Film Restoration does perform much better on that sample.
Upscaling
We have looked into severs upscaling methods using AI technologies. See some results here.
The original 4K image was downscaled to 2K and an AI based method has been chosen to upscale it again to 4K. This images shows an side by side result of that. As you can see the scratches become more prominent since lines got enhanced by that method. Also the upscaler did not deal very good with the grain structure. For this sample a coventional upscaler did perform better.
For this sample the AI based method did perform very well. The sharpness of the caption is remarkable.
We made some good results with the new AI based methods but also some results could not compete with conventional methods. But it is certainly worth to look more closely into AI based methods and we are pretty sure that the new AI based methods will strengthen our restoration filters inside DIAMANT-Film Suite from the year 2020 onwards.
