Research On Automatic Portrait Matting Algorithm Based On Deep Learning

As the basis of image synthesis,image matting is one of the hot issues in the field of image processing.As an important part of image matting,portrait matting technology has a broad application prospect in special effect post production,virtual reality,television creation and other fields.However,most portrait matting methods have problems such as low precision,slow matting speed,and dependence on user assistance,which can not meet the demand for high precision in these fields.To solve the above problems,this paper proposes a lightweight human image matting algorithm based on U-shaped network from the perspective of balancing the precision of matting and the size of the model,which realizes fast and high-precision matting;The segmentation network is used to obtain prior knowledge,and an automatic human image matting algorithm based on the segmentation network is proposed to achieve automatic human image matting without user assistance.The main research contents of this paper are as follows:Firstly,we study the deep learning based matting technology.After an overview of convolution neural network and subsequent convolution operations,Adam algorithm is finally selected for optimization through comparative analysis of the two network optimization algorithms,and network normalization is used to optimize the training process of the model.Then,starting from the principle of matting,we analyze the interaction mode to show the significance of automatic matting.Through the research on the depth learning based matting algorithm,we lay the foundation for the subsequent implementation of matting.Secondly,the construction of portrait dataset and experimental preparation are carried out.In order to ensure the accuracy of training samples,the collected portrait data are screened,and the alpha map with poor labeling effect is finely labeled.Aiming at the problem of insufficient data set samples,image synthesis is used to expand,and online data enhancement is conducted to reduce the sensitivity of the model to images.Because we need to learn the information of the tripartite map in advance during training,we made the tripartite map to complete the experimental preparation.Thirdly,a lightweight human image matting algorithm based on U-shaped network is proposed.Through the improved network and refined network based on UNet,the matting task is completed.The improved network based on UNet retains the U-shaped structure,combines Mobile Net V2 to make the network lightweight,introduces the hswish activation function,and reduces the model calculation cost.Then,the refined network is added to fine tune the obtained alpha graph.In order to avoid adding too many model parameters,the depth separable convolution is used to improve the network expression ability.The experimental results show that the absolute error,SAD and MSE of the improved algorithm are reduced,and the model is greatly reduced,realizing fast and high-precision matting.Finally,an automatic human image matting algorithm based on segmentation network is proposed.In order to solve the problem of relying on manual input of tri graph,the improved network of UNet is used to realize automatic matting.The improved network based on UNet is taken as a pre segmented network branch and the pyramid pool of empty space is added to aggregate the context information to obtain the tripartite graph;Then,the CBAM attention module is embedded in the alpha matting network branch to obtain the multi-scale information of the image and realize the preliminary prediction of the alpha image;Finally,feature fusion is carried out through detail fusion network branches to obtain the final alpha graph.The experimental results show that compared with the DIM algorithm,the SAD and MSE of the algorithm in this paper are reduced by7.6% and 19.4%,respectively.The algorithm realizes automatic matting,which is of great significance for the practicality of human image matting technology.