Virtual Bokeh Reconstruction from a Single System of Lenses

Abstract

We propose a method of virtual bokeh reconstruction by using only multiple differently-focused images acquired with a single system of lenses. First, we analyzed our previously proposed 3-D blurring filter that combines scene information and the acquired multiple differently-focused images to decompose the filter into ray-sets, each of which goes through a certain point on the lens. A new 2-D filter based on dimension reduction was then designed to allow the ray-sets to be reconstructed without the need to estimate the scene. A weighted composition of them provides free shapes of bokeh, i.e., the desired features of lenses and irises. Virtual bokeh reconstruction was simulated by utilizing synthetic multiple differently-focused images with a number of shapes of bokeh. The criteria for appropriate bokeh for image acquisition were also clarified. Finally, we show experimental results of virtual bokeh reconstruction from real images.