Research On Natural Image Matting Algorithm Based On Surrogate Model

As a key technology in digital image processing,including image composition,background replacement,image transfer,and character swaps,image matting has been widely applied in the fields of image editing,computer vision,and virtual reality.The pixel pair optimization of matting has problems such as complex evaluation process and large search space,which makes it difficult to obtain high-quality matting results under low computing resources.This article aims to explore new methods for image matting by focusing on the pixel pairs optimization under conditions of low computational resources,drawing on the idea of surrogate models.Specifically,this study investigates the following research topics:(1)Aiming at the problem that the image matting method based on pixel pair optimization is difficult to provide high-quality foreground alpha matte under low computing resource conditions,from the perspective of the surrogate model,this paper proposes a natural image matting method based on the Gaussian process surrogate model.This method constructs a surrogate model for pixel pair optimization to approximate the pixel pair function,and takes the optimal solution of the Gaussian process surrogate model as a proxy for the optimal solution of the pixel pair function,thus solving the pixel pair optimization problem under a low resource context.Experimental results show that under sufficient computing resources,the image matting method based on the Gaussian process surrogate model can provide high-quality foreground alpha matte results with 1% of the computational resources.(2)The Gaussian process model constructed for random pixel pair sets is difficult to approximate the pixel pair evaluation function when there are few high-quality pixels,which makes it difficult to estimate the high-quality alpha matte problem for natural image matting,from the perspective of pixel sampling,this study proposes a Gaussian Process surrogate model for image matting based on multi-criteria sampling.This method uses multiple distance-based pixel sampling techniques to obtain high-quality candidate pixel subsets from the perspective of global and local similarity,thereby solving the problem of easy pixel loss.Additionally,based on the multiple distancebased pixel sampling strategy,a selection strategy for pixel pair selection based on multiple valuation criteria is designed,which combines multiple pixel pair valuation functions to avoid problems where the solution estimated by a single valuation function is not the optimal solution.Experimental results demonstrate that the proposed Gaussian process surrogate model based on multi-criteria sampling provides a highquality foreground alpha matte and outperforms typical image matting algorithms in cases where high-quality pixels are less abundant.