Natural Background Leaf Image Segmentation Based On Random Walk

Image segmentation is a key step in image processing.It can be described as a process in which images are segmented into a series of connected homogeneous regions according to the similarity criteria and objects of interest are extracted from the background environment by using the low-level visual features of images.Due to the complexity of image segmentation application scenes and the lack of deep cognition of human visual system,there is no method that can be applied to segmentation tasks in all scenes.Therefore,the research of image segmentation method used in specific sceces is significant.Random walk is an interactive segmentation method and has good segmentation performance,which is usually used in image segmentation tasks under complicated backgrounds.In this thesis,we make improvement of random walk algorithm and apply it to leaf segmentation under natural background.The main work of this paper is summarized as follows:In this thesis,the existing random walk method are studied and the shortcomings of these method in leaf image segmentation are analyzed through experiments.Random walk is a typical segmentation method based on graph theory.In order to analyze the random walk,applications of graph theory in image segmentation are introduced.Because graph is directly used in image segmentation,graph construction is important.Some graph models and relevant application are introduced.Several random walk models are introduced and the improvement of these models are studied from graph model point of view.A method of natural background leaf segmentation based on random walk and superpixel is proposed.It can effectively realize the segmentation of leaf image through combining random walk and superpixel.The watershed transform based on multi-scale morphological gradient reconstruction can obtain few superpixels,which have better local spatial information.A method is proposed and estimate the priori information of pixels and superpixels.The superpixels are used to construct the graph together with the pixel layer and label prior.Finally,random walk is used to segment.The unlabled pixels are classified with the maximum probability of first reaching the seed points.The performance of the proposed method is analyzed and discussed through experiments.The proposed algorithm is applied to natural background leaf segmentation,and the segmentation results are compared with some common segmentation methods.The effectiveness of the proposed method is verified from qualitative and quantitative perspectives.The proposed method can achieve better results compared with other methods in same condition.The results are evaluated by three segmentation evaluation indexes and the results of the proposed method can obtain better segmentation evaluation values.