Research On Automatic Keying Method For Non-controlled Environment

In recent years,AR technology that can bring people a new interactive experience has grown rapidly.As one of the carriers for applying AR technology,the virtual studio is also developing rapidly.Traditional live studios set up real scenes in the venue and arrange different program scenes according to needs,which will consume a lot of manpower and material resources.The virtual studio with AR technology can realize the construction and conversion of infinite scenes in a limited space,with low production cost and convenient use.The virtual scene synthesis technology in the virtual studio is to automatically key,generate the background,and synthesize the background through the algorithm,which can provide high-quality and effective AR scenes for the production of programs.Compared with traditional studios,virtual studios can greatly meet people's visual needs.After years of development,the current virtual studio technology has not only greatly improved the key technologies,but also improved the system types,expanded the application fields,and enriched the applications of the virtual studio.The first step in virtual scene synthesis is the keying technology.This paper proposes an automatic keying algorithm,"AIR" automatic keying algorithm,which realizes automatic,constant illumination and real-time in a non-controlled environment.The keying process of the traditional green screen keying algorithm relies on manual and equipment assistance,and the cost is high.In the automatic keying method,the keying algorithm based on the deep learning model is characterized by high robustness,but the keying speed is slower,while the algorithm using the shallow model has a faster keying speed.The "AIR" algorithm proposed in this paper is a simple and effective automatic keying algorithm that does not require additional equipment assistance,greatly reducing the cost of the studio,and while improving real-time performance,it can also take into account the quality of keying.This article proposes a new idea of combining the rough output of the deep learning model with the linear model,and builds a keying framework that combines the deep learning model with the linear model.While solving the contradiction between real-time performance and keying quality,the "AIR" algorithm also breaks through the limitation of traditional live broadcast rooms that require good lighting conditions.It also has a better keying effect in an uncontrolled environment(natural lighting conditions).Based on the original data set,this paper designed and generated a larger data set—Green-2018.This data set has more differences in foreground types and background textures,and more comprehensive evaluation of the method in this paper.This article verifies and analyzes the effect of the "AIR" algorithm on the Green-2018 data set.Experiments show that the "AIR" keying algorithm has a keying effect that is not weaker than that of manual keying software.It is better than the three automatic keying methods compared in this article in terms of accuracy and keying efficiency,and has great advantages in real-time..Experimental analysis further verifies that the "AIR" green screen keying algorithm has excellent keying performance and can perform well in the application scenarios of the virtual studio.