Apprentissage supervisé de regroupement hiérarchique

Image segmentation is a central task in computer vision, providing a simpler representation of an image that focuses on structure. Recently, this fruitful field has been greatly helped by the advent of deep neural networks, which improve performance.However, it remains limited by its single scale, since the structure of things depends on the scale of observation. Hierarchical segmentation provides a multiscale approach to partitioning images into meaningful regions. The scope of the work presented here is the supervised learning of hierarchical image segmentation, namely the training of models (mostly deep neural networks) using images annotated with ground truth hierarchical image segmentations.Achieving such a task requires several prerequisites. First, supervised learning requires annotated data in the desired form, which may not exist. Second, we need a model that can process images, learn from ground truth annotations, and correctly output a hierarchical segmentation that has a particular structure.Third, we need metrics to evaluate the results and compare different hierarchical segmentation methods.This thesis addresses these requirements.In the work presented here, we provide two ways to build hierarchical segmentation datasets, either from scratch by artificially generating them, or by deriving existing object and part segmentation datasets. Several metrics for evaluating hierarchical segmentation are presented, with two main criteria: boundary location and hierarchical order preservation. Several models and loss functions are proposed and compared, starting with a baseline before moving to more complex models focusing on edge precision, crispness, and end-to-end learning