Diffusion-based inpainting methods

Cilj postupaka za popunjavanje nedostajućih dijelova slike (eng. image inpainting) je upotpuniti nedostajući dio slike drugim sadržajem tako da promatrač ne uoči promjenu na slici. Za to postoje različite metode, a u ovom radu podijeljene su na: tradicionalne metode i moderne metode utemeljene na dubokom učenju. Za svaku od metoda opisane su podvrste, a poseban naglasak stavljen je na metode utemeljene na difuziji koje nedostajući dio slike popunjavaju širenjem lokalnih značajki slike iz poznatog dijela slike. Uspoređene su odabrane difuzijske tehnike: transportna, Navier-Stokesova i Fast Marching koristeći različite ispitne slike i maske, a kvaliteta dobivenih slika procijenjena je objektivnim mjerama (PSNR i SSIM).The aim of image inpainting is to complete the missing part of the image with other content so that the observer does not notice the change in the image. There are different methods for inpainting, and in this paper they are divided into traditional methods and modern methods based on deep learning. Subtypes are described for each of those methods, and particular emphasis is placed on diffusion-based methods that fill in the missing region of the image by propagating local image features from a known part of the image. Chosen diffusion techniques (Transport, Navier-Stokes and Fast Marching) were compared using different test images and masks, and the quality of the obtained images was evaluated by objective measures (PSNR and SSIM)