Research On Deep Image Matting Method Based On Image Inpainting And Local Complexity Differences

Image matting technology is widely used in image synthesis,movie production,virtual reality,webcasting,personal video production,graphic and webpage production.As the scene becomes more and more complex,the traditional blue screen matting method is no longer applicable.The research on the problem of natural image matting has gained more and more attention from researchers and communities.This paper uses deep learning tools to solve the problem of natural image matting from the perspective of image inpainting.The main research work and innovations in this paper are as follows:(1)For the current process of matting based on image inpainting,when the trimap is coarse,it will generate more redundant information in the foreground and background inpainting modules.This paper solves this problem by adding a trimap adaptation module to the matting process.The innovation of this research work is: through analyzing the process of foreground and background inpainting,it's recognized that when the coarseness of the trimap gradually increases,the accuracy of the matting will drop sharply.Therefore,this paper designs a new matting process,based on the image inpainting matting method,to improve the matting accuracy when the trimap is coarse.In the experimental part,the composition-1k data set is used to evaluate the effectiveness of the method for improving the accuracy of the matting when the trimap is coarse,and the effect of the method on the coarseness of the trimap is analyzed.(2)Considering the above work,the alpha matte prediction module uses the pixel-level L1 distance loss function,which can't mine the difficult areas in the sample where the complexity of the input image and the real alpha matte are inconsistent for targeted training.Therefore,this paper designs a difficult example mining strategy from the perspective of associated region in the distance loss function.It uses the local complexity-related features of the image to mine difficult areas in the training set,and further improves the matting accuracy.The innovation of this research work is to design an indicator that measures the local complexity difference of the difficult regions and attach it to the distance loss function,so as to do suitable training at different levels for different regions.We verify the effectiveness of this method in improving the matting accuracy through experiments.Finally,combined with the entire work,we conducted a comparative experiment on the public data set with the mainstream matting method.Through comparison and analysis with multiple algorithms,the algorithm proposed in this paper performs best in matting accuracy and details.