Research On Image Matting Algorithm Based On The Combination Of Multi-sampling And Spatial Affine Model

The goal of digital matting research is to accurately separate the foreground in the image from the background.This technology is widely used in image processing,video processing,virtual reality and other fields.The changeable shape of the foreground object,the complex texture of the background area,and the complicated border between the foreground and the background make the matting very complicated.Due to the unconstrained nature of the matting problem,users are required to manually draw the following areas(foreground area,background area and unknown area)to obtain more information for the matting.How to ensure the accuracy of the matting algorithm in a complex environment requires further research.This article analyzes the current problems of natural image matting in depth,and formulates a set of image matting processes under general background.The innovation of this article is based on the following three aspects,First,for the problem of image matting under trimap,many algorithms currently collect sample points from known areas and ignore mining from unknown areas.According to statistics,at least 50%of the pixels in the unknown area are absolute pixels.Aiming at the problem of the limited number of sample points available for collection in the known area,this paper proposes a learning-based two-way dynamic threshold method to mine out some absolute front scenic spots and absolute background points in the unknown area to solve the problem of insufficient sample points,problem.Experimental results show that the preprocessing algorithm can well divide the absolute pixels in the unknown area,and the missing division rate is low.Second,in the current sampling algorithms,when collecting samples,the same sampling method is often used to collect the same number of sample points,ignoring the difference between the foreground and the background.This paper puts forward the concept of simple scene and complex scene,which strengthens the difference between foreground and background.This paper proposes a multi-sampling method.In simple scenes,a parameterized description method is used for the collected sample points,and a point-by-point description method is used in complex scenes.The experimental results show that the value obtained by using the multiple sampling method provided in this article is more accurate.Third,in the previous affine post-processing algorithms,the Closed Form algorithm based on spatial features and color linear models is used to perform post-processing.This color linear model is prone to fail in areas with complex colors.According to the advantages of the KNN method to search for a wide range of samples,a KNN-based post-processing algorithm is proposed to perform initial post-processing.Experimental results show that this post-processing method not only enhances the visual effects of,but also improves the results of,especially the accuracy of the results of the entity foreground and background pixels.