Abstract

Digital image matting is a process of retrieving foreground objects from an image. It has become a widely used technique in Film industry where special effects and editing of an image is actually needed at low cost. In this paper, we propose a novel deep learning based algorithm that can deal with complicated textures and low level features of an image. Our deep model has two basic functions. The first function is a deep convolutional encoder-decoder network that takes an image and trimap values of an image as inputs and finds the corresponding alpha matte for an image. The second function is a small convolutional network that refines the alpha matte predictions of the first network that have more accurate alpha values and sharper edges. Experimental results with 1000 training and testing images show that the proposed model drastically increased the accuracy level over existing models.

Paper Details