Tree image extraction, which is more than basic, provides the fundamental data and technical support for these studies. However, tree image extraction is also a difficult problem since the tree photo shooting in the natural scene contains a wealth of information and is subject to the environment, light, weather, noise and other interference. The diversity of trees and its surrounding scenery make tree image extraction in the natural scene complex and challenging. This paper adopts the natural image matting to extract the tree image, better solve the problems of tree image internal void and image transparent, propose the concept of regions of interest, and introduce the regiongrowing method to improve the matting based on Markov Random Field (MRF) from the aspects of simplifying the division of trimap, determining the foreground pixels as much as possible and reducing the number of pixels of the unknown region. The experiment shows that this method can effectively extract the tree image, simplify the human machine interaction process, and enhance color accuracy, and at the same time increase computing speed greatly.