This research project was made under my internship at Institute for Information Transmission Problems of the Russian Academy of Sciences (Kharkevich Institute) (http://iitp.ru/en/science) and is based on the article https://arxiv.org/abs/1805.01934
The problem solved is adjusting the existing NN pipeline for an arbitrary photo camera with the aim to transform short exposed raw photos into high-quality RGB images with the help of U-net. PSNR and SSIM metrics are used to compare output of the NN with the ground truth (long exposed photo). The problem of applicability of a pretrained model for a new camera as well as the reducing of the train set size are investigated.
Code snippets for training and testing are presented above in Tensorflow 2.0.