L. Khelifi
Ph.D. thesis, University of Montreal
Publication year: 2017

 

ABSTRACT

This thesis is dedicated to study two complementary problems, namely the fusion of image segmentation and the semantic interpretation of images. Indeed, at first we propose a set of algorithmic tools to improve the final result of the operation of the fusion. Image segmentation is a common preprocessing step which aims to simplify the image representation into significant and spatially coherent regions (also known as segments or super-pixels) with similar attributes (such as coherent parts of objects or the background). To this end, we propose a new fusion method of segmentation in the sense of the Global consistency error (GCE) criterion. GCE is an interesting metric of perception that takes into account the multiscale nature of any segmentations of the image while measuring the extent to which one segmentation map can be viewed as a refinement of another segmentation. Secondly, we present two new approaches for merging multiple segmentations within the framework of multiple criteria based on a very important concept of combinatorial optimization ; the multi-objective optimization. Indeed, this method of resolution which aims to optimize several objectives concurrently has met with great success in many other fields. Thirdly, to better and automatically understand the various classes of a segmented image we propose an original and reliable approach based on an energy-based model which allows us to deduce the most likely classes by using a set of identically partitioned segmentations (in the sense of a certain criterion) extracted from a learning database (with pre-interpreted classes) and a set of semantic likelihood (energy) terms