The Semantic Object Segmentation Method Research

Generally, there are multiple objects in one image, and we are only interested in the objects which can express the semantic information of the image. This has required us extracting the semantic objects, resulting in the coming of the semantic object segmentation technology. Nowadays, the technology has become an important branch of image segmentation, and been widely used in the fields such as image editing, video coding and video surveillance.Traditional semantic object segmentation methods first train a classifier by using the image examples, detect the semantic objects with the classifier, and finally segment the semantic objects based on the detection results. Since the segmentation is related to the detection, the results obtained by these methods are affected by the performance of the classifier. In contrast, the popular semantic object segmentation methods try to find the underlying semantic informations by user interactions or the inherent properties of the images, utilize the informations to guide the segmentation. Since these methods perform in a manner like supervision, better results can be obtained.This paper focuses on the popular semantic object segmentation methods, i.e., image cosegmentation and the segmentation of the low depth-of-field images. By studying and analyzing the graph-based image segmentation method, we propose two semantic object segmentation methods as follows:1. Based on the formulation of Ncut, a spectral cosegmentation method is proposed. In the method, we first build the graph model for the input image pair and construct the cost function. Then the knowledge of matrix is introduced to turn the the cost function minimization problem into the eigen-decomposition problem. Finally, spectral clustering is utilized to segment the common objects of the images.2. Based on graph cut and the analysis of the amplitude spectrum, a method for segmenting the low depth-of-field images is proposed. In the method, we first analyze the amplitude spectrum of the low depth-of-field images, so as to obtain the amplitude decomposition model. Then we compute the focus maps which can detect the focused objects. Finally, we binarize the focus maps to get the initial segmentation result, and utilize graph cut to refine the boundary.The experiments show that the methods proposed by us can segment the semantic objects accurately and effectively.