This project deals with an interdisciplinary domain which includes image processing algorithms, linear programming problems. The project focuses on the “Photometric Copy Paste” algorithm
Abstract
This project deals with an interdisciplinary domain which includes image processing algorithms, linear programming problems and C++ programing. The project focuses on the “Photometric Copy Paste” algorithm which was developed in The HP research labs and its goal is to implement a real-time application that recolors a picture (or part of it) in the “foot steps” of another picture. The need for this application stems from a lack in a product with high image processing abilities, which enables performing complex manipulations on pictures, and is easy to use by a non-expert user.
The algorithm has two main parts: The first one focuses in detecting the source and target regions (who are we and who do we want to look like) with a segmentation algorithm based on a Bayesian classifier. The second one, which is the core of the entire “Photometric Copy Paste” algorithm is based on the solution of a linear programming problem, a variation of the transportation problem. The main idea is transporting a “probablity mass” of colors from one region to another, and the target function of the linear programming problem is to bring to a minimum the “inter-color distance”, or in lamen’s terms, make the colors in one region be as similar as possible to colors in another region.
Working on the project included implementing the suggested algorithm in C++, with the aim of reaching satisfying real time performance.
The implementation itself did not end in the desired results. Moreover, the simulation code which was supplied in the beginning of the project was found to produce unexpected outputs for certain inputs.
Results
simulation input
simulation output
