Efficiency evaluation of background removal algorithms

Abstract

The necessity to delete background on an image occurs quite often. Technologies are closely connected to our everyday life so with its change they have to be adapted to the new realities. The increasing amount of video conferences during the COVID19 epidemic brought the concept of virtual backgrounds to the fore. Any videoconference services either Google Meet, Zoom or MS Teams and others have this function. Virtual background could be really useful for confidential protection or hiding one's chaotic surroundings. The technology of automated background removal must be developed, because first of all, people want to leave their personal life private, that is why they hide the place where they are. Secondly, the usage of this technology is applied for visual tracing and image segmentation. But more often background removal is used when making new content, it is possible to add any kind of background to an extracted fragment thereby getting a completely new image. Currently, there are a lot of programs with a help of which a background removal could be done, but manual correction still remains the most trusted method, yet it is a labourintensive process with requires a lot of time. Frequently people need a quick solution for image editing and the services that provide such features are either paid services or they are not able to cope with the task, presenting the necessary level of quality. The article explores different modules of background removal and evaluates the efficiency of their work. The actuality of this problem is caused with the need of confidentiality protection with the increasing usage of services for videoconferences and also for making content, visual detection and operative problem solution. In this material we take a look at the methods used in the models, evaluate the work of the most popular models for background removal with the help of proposed new metrics for output results. All of this helps to define pluses and minuses of studied algorithms.