Stacked Deep Multi-Scale Hierarchical Network for Fast Bokeh Effect Rendering from a Single Image

The Bokeh Effect is one of the most desirable effects in photography for\nrendering artistic and aesthetic photos. Usually, it requires a DSLR camera\nwith different aperture and shutter settings and certain photography skills to\ngenerate this effect. In smartphones, computational methods and additional\nsensors are used to overcome the physical lens and sensor limitations to\nachieve such effect. Most of the existing methods utilized additional sensor's\ndata or pretrained network for fine depth estimation of the scene and sometimes\nuse portrait segmentation pretrained network module to segment salient objects\nin the image. Because of these reasons, networks have many parameters, become\nruntime intensive and unable to run in mid-range devices. In this paper, we\nused an end-to-end Deep Multi-Scale Hierarchical Network (DMSHN) model for\ndirect Bokeh effect rendering of images captured from the monocular camera. To\nfurther improve the perceptual quality of such effect, a stacked model\nconsisting of two DMSHN modules is also proposed. Our model does not rely on\nany pretrained network module for Monocular Depth Estimation or Saliency\nDetection, thus significantly reducing the size of model and run time. Stacked\nDMSHN achieves state-of-the-art results on a large scale EBB! dataset with\naround 6x less runtime compared to the current state-of-the-art model in\nprocessing HD quality images.\n