Research On Image Matting Method Based On Dual Affine And Shared Sampling Criterion

Digital image matting is a classic problem in image processing.At first,the digital image matting technology is usually blue screen matting.Because of its single background color,this kind of matting is relatively simple,but this kind of blue screen matting has higher requirements for the background of the image.At present,the more popular method is the image matting method under the general background,that is,natural image matting.This method needs to input a trimap template to preliminarily specify a part of the opaque foreground,the opaque background region.Natural image matting is an image processing technology that accurately separates the foreground information and background information of the input image.At the same time,the foreground information can be combined into the background of another image through image synthesis technology.At present,natural image matting still has many problems.Mainly for the following reasons:It is difficult to accurately calculate the probability that the pixel in this region belongs to the foreground pixel for the hair,foreground edges,and regions with high transparency of the image.Only under the condition of good manual labeling can obtain a satisfactory matting result.Therefore,this paper mainly proposes three innovative methods to address the limitation of the range of the search for natural image matting,the large error of known pixels division,and the lack of a distinct sense of hierarchy:multi-scale space fusion method based on affine criteria,the shared sampling method combined with nonlocal affine criterion and reconstructed alpha image matting method based on convolution fusion.(1)Aiming at the problems existing in the method of affine image matting,the search range is very fixed,the search range cannot be changed according to the actual input of image trimap and the visual effect is not good,this paper proposes a multi-scale spatial fusion method based on the affine criterion.The method first using static threshold method WCT algorithm to preprocess the image.Secondly,a two-level spatial search method is adopted,so that different trimaps input can cover more search ranges,and further matting calculations are performed on the local details of the image based on the assumption of the local smoothness of the foreground and background colors.Finally,a sparse linear equation is solved to calculate the final transparency alpha result.The experimental results show that,compared with the traditional affine image matting method,the alpha value obtained by the method proposed in this paper is more accurate and the visual effect is better.(2)Aiming at the problems existing in the method of sampling image matting,such as the wrong division of the known region,unclear sense of hierarchy and large error value of the unknown region solution,this paper proposes the shared sampling method combined with nonlocal affine criterion.Firstly,the iterative static threshold method is adopted to preprocess the image,and the trimap image is preliminarily extended to obtain some opaque foreground pixels and opaque background pixels.Secondly,a shared sampling method based on color similarity is adopted to sample pixels with similar features through information transmission among neighborhood pixels,and perform sample refinement and local smoothing.Finally,Post-processing of the alpha image obtained from the sampling using a nonlocal affine matting algorithm.The experimental results show that,compared to only using a single color sharing method,the method proposed in this paper can more accurately divide opaque pixels and have a more distinct sense of hierarchy,which can not only enhance the visual effect of the final alpha image,but also improve the accuracy of the alpha value.(3)Aiming at the cooperative relationship between local and nonlocal matting methods in traditional methods,a reconstruction alpha image matting method based on convolution fusion was proposed.In this method,the initial alpha image and normalized RGB color image obtained by the previous two traditional methods are used as input,and the end-to-end mapping between the input and reconstructed alpha matting is directly learned.This method uses the traditional affine and sampling methods proposed in Chapter 2 and Chapter 3 as two algorithms to generate the initial alpha images.They complement each other in local and nonlocal principles.The main advantage of this method is that it can identify different local image structures and then effectively combine local and nonlocal matting results to obtain higher quality alpha results.