BiLEARN
Bilevel learning for joint motion and image reconstruction
Characterizing the dynamics involved in the data processing of images provided by novel high resolution technologies is a key task in many applied sciences, ranging from physics to medicine. Essential steps in this direction are to identify velocity fields between subsequent image frames, as well as to cope with the presence of possible low resolution and noise effects in the single frames. A crucial difficulty is thus to tune reconstruction parameters in such a way to guarantee the best image and motion reconstruction at the lowest computational cost.
In this subproject, we plan to develop bilevel learning schemes for simultaneous motion reconstruction and image denoising, as well as for the case in which motion reconstruction is combined with segmentation effects.
The innovative character of this subproject is twofold. First, we will initiate the study of bilevel learning schemes involving also motion reconstruction. This will go far beyond the current variational theory for classical static problems. Second, allowing for fractional effects and extending the well-posedness analysis available so far for joint motion estimation and image reconstruction to dynamic inpainting, will enhance fine texture details which will be relevant in medical applications.