Texture Missing Parts Generation Based on Image Statistical Analysis
Artem Melnychenko, Y. Vodolazskyi · Cybernetics and computer engineering · 2023
Introduction.Restoration of damaged images is a long lasting problem that currently does not have a generalized solution.Many methods which are being used nowadays are damage type specific, which means that for each case of damaged image an algorithm must be picked by a human.A state of the art generative algorithms, which may handle many of the damage types, still lack the precision and require huge training datasets.Thus an algorithm that is able to handle most common damage types and does not demand lots of time and computational power is still in need.The purpose of the paper is to research the current state of the art algorithms that solve texture missing part generation problem as well as to propose a new method, which might provide both precision and ease of use for solving said problem for most of the damage types using the same approach. Methods.Research and analytics are used for processing found literature on the topic to substantiate the main approaches and best practices for the solution of the texture missing parts generation problem.As for purposed method, Gibbs sampling is used as a means of generating missing pixels of the image.Some additional algorithms, which might be used to generate probabilistic distribution for sampler and the means of getting the pixel value from the sampling process, are mentioned in the article itself.Results.State of the art approaches for solving texture missing parts generation are analyzed and compared.Main groups of generative, texture reparation, gradient filling and combined methods are described and compared.New method for generating missing parts of the texture based on statistical analysis of the scene images is proposed.The generation of the pixel values in said method is based on Gibbs sampling.The first results of purposed method with patch based probabilistic distribution generation are shown.Conclusions.The proposed Gibbs sampling based method is able to provide results, which are comparable with those generated by other modern methods.As a future work, it is planned to develop new more sophisticated and precise patches matching algorithms as well as to research other methods of both generating probability distribution and gathering pixel value from the sampling process.