Why Maximum-A-Posteriori Blind Image Deblurring Works After All
Gian Marti, Boxiao Ma, Hans‐Andrea Loeliger · 2021 29th European Signal Processing Conference (EUSIPCO) · 2021
Maximum-a-posteriori (MAP) methods, while being a standard choice for many estimation problems, have been considered problematic for blind image deblurring: They have been suspected of preferring blurry images to sharp ones. Alternative methods without this apparent defect have been proposed instead. Reservations about MAP methods for blind image deblurring persist even as their close relation to these alternatives has become evident. We revisit the literature on this topic and argue that the original rejection of MAP methods was ill-founded. We show that the MAP approach can prefer sharp images over blurry ones. Furthermore, we show experimentally that the MAP approach can in principle achieve deblurring results that are competitive with the allegedly superior methods. We thereby challenge some traditional notions of the relevant causes underlying successful blind deblurring to obtain a more accurate understanding of the blind image deblurring problem.