Research On Image Segmentation And Scene Understanding Using Discriminative Learning Method

In the field of computer vision, image segmentation is an important but a difficult task. In this paper we propose a new framework that combines the texture features and geometrical features to transform the segmentation problem to a classification problem. According to the different objects in different segmentation problems, we further study the choice of features and discriminative learning functions. Moreover, we discuss further in three typical applications, texture image segmentation, outdoor scene understanding and indoor scene understanding. The main works of this paper are as follow:1. We introduce the framework that combines the texture features and geometrical features, and give the optimization objective that transform the segmentation problem to the classification problem.2. For the texture image segmentation, we further divide it into two sub-problem, interactive texture-based image segmentation and texture primitives based scene segmentation, to discuss. For the interactive texture-based segmentation, we choose histogram of gradient (HoG) features and boosting algorithm for segmentation. For the texture primitives based scene segmentation, we choose textons features and the Joint Boosting algorithm for segmentation. The experimental results show that the learning method based on the texture features and discriminative model, can effectively distinguish between the various texture patterns in the image. Meanwhile, this method proves to be able to recognize and segment regions with different semantic contexts in various kinds of scenes.3. We further discuss the problem of three-dimensional scene understanding. To solving both the outdoor scene understanding problem and indoor scene understanding problems, we propose an algorithm for image understanding that is based on geometric features. The results show that even in the complex scenes, our algorithm still can get reliable results.