Coarse Semantic Guided Alpha Matting via Simultaneous Foreground and Background Estimation

Abstract

To address the "color spill" problem in image matting, an algorithm of simultaneously predicting the foreground, background and alpha matte is proposed, and an automatic matting network combined with coarse semantics is designed. The algorithm method includes two stages: coarse semantics generation and matting via simultaneous foreground and background prediction. In the first stage, a semantic segmentation network is used to obtain intermediate results, and then a coarse semantic information fusion module is applied to roughly estimate the semantic information on multiple scales. In the second stage, an encoder-decoder is used to refine the semantic information and obtain the prediction of foreground, background and alpha matte. Foreground objects can be extracted more accurately through this two-stage network. Also, the obtained alpha matte can be directly used for downstream tasks such as image and video processing. The experimental results on Adobe dataset and the Distinction-646 dataset show that the sum of absolution difference is 42.5 and 50.3, the gradient error is 27.1 and 28.0, respectively. And the details are also more accurate.