Deep Eyebrow Matting

In high-quality 3D face reconstruction,hair such as eyebrows will greatly affect the reconstruction.Eyebrows are often used for the semantic editing of portraits.To remove the eyebrows,or change their shape,position,and color,we need pixel-level precision control of the eyebrows,namely eyebrow extraction,and eyebrow matting.However,the fine structure of the eyebrows and the diversity of the face skin bring difficulties to eyebrow matting.Traditional matting is usually time-consuming and requires skilled professionals to manually process images,a more effective way is to apply automatic image matting for eyebrow extraction.We note that there is no eyebrow matting dataset,and existing image matting methods also cannot generate eyebrow's alpha matte accurately.To tackle the lack of eyebrow matte data,this thesis proposes a method of constructing an eyebrow matting dataset by rendering realistic 3D eyebrows and human faces.Our method generates the eyebrow data and the corresponding accurate alpha matte respectively by rendering the 3D models of the eyebrows under different backgrounds.In order to solve the occlusion problem of the character model,we render the eyebrow segmentation results including the occlusion relationship and use the segmentation result as a mask to obtain an accurate alpha matte.Our method trains an end-to-end automatic eyebrow matting model on the constructed rendering matting dataset.Meanwhile,this paper proposes two methods to solve the problem of different data distribution between the rendered image and the real image.The first is a progressive training strategy.By adding high-confidence results to the training set,after several iterations,the model can quickly adapt to the difference between the rendered image and the real image.In addition,this paper introduces a novel domain adaptation method in image matting,in order to improve the matting quality of the model for the real images domain.This method first inputs the rendered image and the real image in pairs to the encoder to extract the features,then inputs the features into the discriminator,sets the adversarial discrimination loss,and finally forces the encoder to extract more closer features in two fields.As a result,the network can adapt to the differences between different domains.To validate the advantage of our method,we present a new eyebrow matting test dataset to test the effectiveness of each step of the iterative process.We qualitatively and quantitatively compare our method to Deep Image Matting and other state-of-the-art methods.The results show that our method achieves the best eyebrow matting performance.