Constrained Predictive Filters for Single Image Bokeh Rendering

Bokeh rendering is a technique used to take pictures with out-of-focus areas to highlight regions of interest. Due to limitations in hardware and shooting condition, rendering a bokeh image from a full-focus image has attracted a lot of interest. In this paper, we model bokeh rendering as the combination of salient region retention and bokeh blurring, and propose a neural network to generate a realistic bokeh image from a single full-focus image through end-to-end training. Specifically, we propose a gate fusion block to estimate the salient area, and introduce a constrained predictive filter for salient region retention and bokeh blurring within a unified architecture. Further, we utilize a pixel coordinate-based map to enhance the training. Experimental results illustrate the effectiveness of our model. The comparison with state-of-the-art methods (PyNET [18], DMSHN [12], BGGAN [40], etc.) shows that our model produces better bokeh effects and retains salient objects.