A Learning Approach for Adaptive Image Segmentation

Abstract

In this chapter, we have proposed a learning approach for three major issues of image segmentation: context adaptation, algorithm selection and parameter tuning according to the image content and the application need. This supervised learning approach relies on hand-labelled samples. The learning process is guided by the goal of the segmentation and therefore makes the approach reliable for a broad range of applications. The user effort is restrained compared to other supervised methods since it does not require image processing skills: the user has just to click into regions to assign labels, he/she never interacts with algorithm parameters. For the figure-ground segmentation task in video application, this annotation task is even automatic.