Informative image partitioning: techniques and applications

Abstract

It is natural for us as humans to pay more attention to some parts of what we see in the world around us than others. This thesis introduces a technique for partitioning an image in such a way that 'interesting' objects are segmented into fine segments while 'uninteresting' objects are segmented into coarse segments, called an Informa- tive Image Partitioning. This is achieved by first producing an oversegmentation of an image, dividing the whole image into small segments. These segments are then merged together in uninteresting areas to create different scales of segmentation in different parts of the image. Two different techniques are presented for producing an Informative Image Part- itioning from user input. The first is based on learning a distance metric and a threshold to determine if two adjacent oversegments should be merged. The second method introduces a novel edge to node transformation and trains a binary Support Vector Machine classifier on this new graph to determine if a pair of oversegments should be merged. As the translation from the edge graph to an image partitioning is not one to one, three loop resolution methods are introduced to resolve potential ambiguities. These techniques are able to generate an image partitioning which has the desired 'focussing' behaviour. A key feature of the techniques detailed in this thesis is that the 'interestingness' of different objects is defined by the user using simple side information, and objects do not need to be explicitly classified in order to define them as interesting or uninteresting. This thesis also explores the behaviour of these methods. It shows that it is pos- sible to train a task-specific model for generating an Informative Image Partitioning on similar images. In addition it demonstrates the impact of different user input on the partitioning developed, showing that different objects are able to be defined as the object of interest in an image. Further, the thesis details the performance of the developed methods using only a single channel of information. Two applications of Informative Image Partitioning, image and video compres- sion and stereo disparity calculation, are also discussed. When used as a part of an image compression algorithm, high image and video compression results are achieved, and compressed images are both qualitatively and quantitatively good. The Informative Image Partitioning process can also be used to improve stereo dis- parity by making use of a piecewise planar assumption and treating planar objects as uninteresting. By treating each partition within a processed image as a plane, it is possible to generate disparity maps with better performance in large, low informa- tion areas in which disparity calculation methods have traditionally had difficulty. Show more