Intelligent Editing And Synthesis Of Highly Realistic Visual Media

With the development of computer technology and Internet, visual media has become an important medium of display and transmission. Compared with other medium,visual media is digital, universal and intuitive. With the development of the visual media collection devices, high quality visual media has become an indispensable way to carry,store and display for users. Meanwhile, with the development of hardware, editing and synthesis of visual media is widely used in various commercial software and industry areas, such as advertising, news, movies, and mobile applications. Since users can access to more high quality visual media, editing and synthesis of visual media is not limited to the availability of results. High intelligence and high reality becomes more important.This paper focuses on editing and synthesis of visual media. To make it more intelligent and realistic, this paper focuses on visual media editing and synthesis, deeply researches edit propagation algorithms, change blindness images, shadow removal problems, and proposes a series of methods. The main contributions are:? We propose a novel antialiased edit propagation algorithm. Traditional methods lack antialias information, thus produce alias artifacts along boundaries. We analysis alias artifacts and propose Antialias Map to describe antialias information. After merging Antialias Map with edit propagation optimization, alias artifacts vanish.? We propose an automatic change blindness image synthesis algorithm with controlled di?culty. Traditionally these images are manually generated, and the level of di?culty is hard to control. Based on cognitive psychology, we deeply research change blindness, and propose the first quantitative metric. Then we develop this algorithm, which improves the intelligence of change blindness image synthesis.? We propose an automatic single-image cast shadow detection and removal algorithm. Cast shadow destroys image structure and limit the usage of detection methods.Traditional methods are less universal and realistic, and rely on user interaction.This newly proposed method is based on patch match framework, uses parameter learning and model training strategies to preserve the robustness of the whole system, thus produces high realistic shadow removal results.