Efficient Bokeh Effect Rendering using Generative Adversarial Network

The bokeh effect in photography is an artistic effect that makes the subject stand out by intentionally distorting the focus of the background while focusing on the subject. In this paper, we propose an end-to-end deep learning model that generates bokeh images from wild HD images. The proposed generative adversarial network model consists of a generator and a double-scale discriminator. It is so lightweight that it can carry out bokeh effect rendering in 0.03 seconds per image on a PC. In addition, we show that in a mobile environment, HD images can be processed in 1.3 seconds per image.

1. Introduction

With the development of information and communication technologies, the use of the Internet and smartphone devices has become common in modern society. Accordingly, consumers can easily and quickly obtain the information they want without restrictions in time and space. In particular, image-based social network services such as Instagram and Pinterest can compress various moments of everyday life through image information. Such visual information is relatively faster and easier to share than written information. Especially, unlike photos that show objects as they are, photos with controlled light, depth, and white balance can effectively convey users' opinions, ideas, emotion, mood, etc. Recently, attempts have been made to implement photo effects using deep learning to make it easier for users to capture their intentions with smartphones instead of highend professional cameras. Among them, bokeh effect rendering is a relatively new research problem that has emerged over the past few years.