Improving Hair Segmentation using Image Matting

Abstract

Many real-life intelligent applications such as human beautification and recommender systems, demand accurate hair segmentation from a single portrait image. Recent advances in deep learning achieve or even surpass human-level performance in a multitude of intelligent applications ranging from image classification to game playing. However, due to the lack in a large scale of facial images labeled with high-quality hair masks, current deep learning models appear to have difficulty classifying hair strands accurately. For this research, focusing on semantic hair segmentation, we investigate the use of image matting in deep learning to generate accurate hair masks. In particular, we apply image matting to expand from an initial segmented hair region from deep learning models, via different sampling and propagation matting methods to recover missing hair strands. Afterwards, an ensemble learning approach is proposed to find an optimal performance through the combination of multiple matting results. The experimental results show that image matting applied to hair masks can achieve notably improved results with fine details in hair segmentation.