fmerizzi / deep_prior

some general notes and approaces on deep imae prior

deep_prior

some general notes and approaces on deep image prior

paper notes

1. the structure of a generator network is sufficient to capture image statistics
2. INVERSE PROBLEMS, randomly initialized network can be used as prior
3. THE STRUCTURE OF THE NETWORK MUST RESONATE WITH THE STRUCTURE OF THE DATA
4. you need a generator network

---

1. I start with some random weights to the NN and iteratively update THEM
2. updating with a comparison to the original image, super resolution is reasonable
3. a network untrained descends much faster towards a "real" image rather than random noise.
4. in inpainting, I'm just calculating the loss on the non-mask pixels! sounds pretty reasonable!
5. in general I should tune the network for every specific image..
6. for LARGE HOLE INPAINTING, an inizialization with meshgrid gradient rather than uniform noise
7. noise based regularization, perturb the input
8. the convolutional operation impose self-similarity on the generated images

future developments

• taking into consideration the likeness of a network to rebiuld random vs non random images. maybe related to a so called random distance..
• what if we sample the random creations a CNN and use them instead of random layers? huh?