Portrait Matting Algorithm Based On Progressive Segmentation Network

Portrait matting is a hot issue in the current image processing direction,which has been applied in many prospects of digital entertainment.The core of portrait matting lies in obtaining high-precision edges of portraits from natural images.Existing portrait matting methods use trimap images(foreground,background,and unknown regions)as a priori information to constrain the segmentation process of foreground portraits and background regions,thereby obtaining high-precision alpha matte of portraits.However,the trimap images of the existing methods require manual labeling,which is difficult to apply to large-scale data or time-limited application scenarios.And manual labeling is subjective,and when the trimaps annotation is inaccurate,the segmentation accuracy of the foreground portrait and background areas is also falling.It is difficult to obtain high-precision alpha matte for portraits.In view of the above problems,this paper uses deep learning segmentation network and attention model to realize the automatic generation of trimap images,and proposes a fullautomatic portrait matting method based on progressive segmentation network.In this method,the trimap image is regarded as the result of coarse segmentation,and the final alpha matte is regarded as the result of fine segmentation.An end-to-end deep network framework of trimap coarse segmentation subnetwork with alpha matte fine segmentation subnetwork is designed.Specifically,the coarse segmentation subnet regards the generation of trimap as a threecategory task of semantic segmentation,uses the residual network Res Net50 as the basic network framework,and combines the spatial attention module to enhance the feature extraction ability of the rough segmentation subnetwork;The fine segmentation subnetwork uses a encoding-decoding network for fitting and predicting the alpha matte,and combines the foreground pixel classification probability and unknown area classification probability of the coarse segmentation subnet to refine the predicted alpha matte to obtain a more accurate alpha matte.The experimental results show that,compared with the method of manually marked trimap,the algorithm in this paper ensures the quality of the trimaps,and the speed of generating the trimaps is greatly improved.The generated speed of the trimap is 13.5fps.Compared with the Deep Image Matting method based on deep learning,the algorithm matting speed of this paper is increased by 0.78 fps,and the matting accuracy SAE and MSE are improved by 2.136 points and 1.797 points respectively.The algorithm in this paper automates the entire process of the portrait matting algorithm.It combines the trimap generation and the matting work into a whole task,which improves the matting efficiency and accuracy.It is of great significance for the practical application of the portrait matting technology.