Priors Good priors help to solve inverse problems that are illposed
due to missing, noisy, or blurry data. Instead of choosing
just one, we have concentrated on three: TV, the denoising / selfsimilarity
prior, and the cross-channel prior. They are very useful,
but they may also introduce artifacts: TV can create unnatural
“watercolor-like” edges; self-similarity helps to reconstruct missing
data, but it only works well if the image actually exhibits selfsimilarity—in
regions with unique patterns, such as vegetation, it
may over-smooth; and while the cross-channel gradient prior helps
to avoid color fringes, it sometimes mistakes colored pixels for ar