Semi-Supervised Learning Matting Algorithm Based on Semantic Consistency of Trimaps

Image matting methods based on deep learning have made tremendous success. However, the success of previous image typically relies a massive amount pixel-level labeled data, which are time-consuming and costly to obtain. This paper first proposes semi-supervised algorithm semantic consistency trimaps (Tri-SSL), uses provide weakly supervised signals for unlabeled reduce labeling cost. Tri-SSL is single-stage that consists branch share same network in one iteration during training. The consistent with standard methods. In branch, different granularities used as images, two naturally perturbed samples. Orientation constraints imposed prediction results granuliarty intermediate features network. Experimental show improves model performance by effectively utilizing data.