Segmentation hiérarchique faiblement supervisée

Image segmentation is the process of partitioning an image into a set of meaningful regions according to some criteria. Hierarchical segmentation has emerged as a major trend in this regard as it favors the emergence of important regions at different scales. On the other hand, many methods allow us to have prior information on the position of structures of interest in the images. In this paper, we present a versatile hierarchical segmentation method that takes into account any prior spatial information and outputs a hierarchical segmentation that emphasizes the contours or regions of interest while preserving the important structures in the image. An application of this method to the weakly-supervised segmentation problem is presented.

Prior-based Hierarchical Segmentation Highlighting Structures of Interest

Image segmentation is the process of partitioning an image into a set of meaningful regions according to some criteria. Hierarchical segmentation has emerged as a major trend in this regard as it favors the emergence of important regions at different scales. On the other hand, many methods allow us to have prior information on the position of structures of interest in the images. In this paper, we present a versatile hierarchical segmentation method that takes into account any prior spatial information and outputs a hierarchical segmentation that emphasizes the contours or regions of interest while preserving the important structures in the image. Several applications are presented that illustrate the method versatility and efficiency.

Automatic Selection of Stochastic Watershed Hierarchies

The segmentation, seen as the association of a partition with an image, is a difficult task. It can be decomposed in two steps: at first, a family of contours associated with a series of nested partitions (or hierarchy) is created and organized, then pertinent contours are extracted. A coarser partition is obtained by merging adjacent regions of a finer partition. The strength of a contour is then measured by the level of the hierarchy for which its two adjacent regions merge. We present an automatic segmentation strategy using a wide range of stochastic watershed hierarchies. For a given set of homogeneous images, our approach selects automatically the best hierarchy and cut level to perform image simplification given an evaluation score. Experimental results illustrate the advantages of our approach on several real-life images datasets.

Flooding edge or node weighted graphs

Reconstruction closings have all properties of a physical flooding of a topographic surface. They are precious for simplifying gradient images or, filling unwanted catchment basins, on which a subsequent watershed transform extracts the targeted objects. Flooding a topographic surface may be modeled as flooding a node weighted graph (TG), with unweighted edges, the node weights representing the ground level. The progression of a flooding may also be modeled on the region adjacency graph (RAG) of a topographic surface. On a RAG each node represents a catchment basin and edges connect neighboring nodes. The edges are weighted by the altitude of the pass point between both adjacent regions. The graph is flooded from sources placed at the marker positions and each node is assigned to the source by which it has been flooded. The level of the flood is represented on the nodes on each type of graphs. The same flooding may thus be modeled on a TG or on a RAG. We characterize all valid floodings on both types of graphs, as they should verify the laws of hydrostatics. We then show that each flooding of a node weighted graph also is a flooding of an edge weighted graph with appropriate edge weights. The highest flooding under a ceiling function may be interpreted as the shortest distance to the root for the ultrametric flooding distance in an augmented graph. The ultrametric distance between two nodes is the minimal altitude of a flooding for which both nodes are flooded. This remark permits to flood edge or node weighted graphs by using shortest path algorithms. It appears that the collection of all lakes of a RAG has the structure of a dendrogram, on which the highest flooding under a ceiling function may be rapidly found.

Watersheds on edge or node weighted graphs "par l'exemple"

Watersheds have been defined both for node and edge weighted graphs. We show that they are identical: for each edge (resp.\ node) weighted graph exists a node (resp. edge) weighted graph with the same minima and catchment basin.

Watersheds, waterfalls, on edge or node weighted graphs

We present an algebraic approach to the watershed adapted to edge or node weighted graphs. Starting with the flooding adjunction, we introduce the flooding graphs, for which node and edge weights may be deduced one from the other. Each node weighted or edge weighted graph may be transformed in a flooding graph, showing that there is no superiority in using one or the other, both being equivalent. We then introduce pruning operators extract subgraphs of increasing steepness. For an increasing steepness, the number of never ascending paths becomes smaller and smaller. This reduces the watershed zone, where catchment basins overlap. A last pruning operator called scissor associates to each node outside the regional minima one and only one edge. The catchment basins of this new graph do not overlap and form a watershed partition. Again, with an increasing steepness, the number of distinct watershed partitions contained in a graph becomes smaller and smaller. Ultimately, for natural image, an infinite steepness leads to a unique solution, as it is not likely that two absolutely identical non ascending paths of infinite steepness connect a node with two distinct minima. It happens that non ascending paths of a given steepness are the geodesics of lexicographic distance functions of a given depth. This permits to extract the watershed partitions as skeletons by zone of influence of the minima for such lexicographic distances. The waterfall hierarchy is obtained by a sequence of operations. The first constructs the minimum spanning forest which spans an initial watershed partition. The contraction of the trees into one node produces a reduced graph which may be submitted to the same treatment. The process is iterated until only one region remains. The union of the edges of all forests produced constitutes a minimum spanning tree of the initial graph.

The steepest watershed: from graphs to images

The watershed is a powerful tool for segmenting objects whose contours appear as crest lines on a gradient image. The watershed transform associates to a topographic surface a partition into catchment basins, defined as attraction zones of a drop of water falling on the relief and following a line of steepest descent. Unfortunately, catchment basins may overlap and do not form a partition. Moreover, current watershed algorithms, being shortsighted, do not correctly estimate the steepness of the downwards trajectories and overestimate the overlapping zones of catchment basins. An arbitrary division of these zones between adjacent catchment basin results in a poor localization of the contours. We propose an algorithm without myopia, which considers the total length of a trajectory for estimating its steepness. We first consider topographic surfaces defined on node weighted graphs. The graphs are pruned in order to eliminate all downwards trajectories which are not the steepest. An iterative algorithm with simple neighborhood operations performs the pruning and constructs the catchment basins. The algorithm is then adapted to gray tone images. The graph structure itself is encoded as an image thanks to the fixed neighborhood structure of grids. A pair of adaptative erosions and dilations prune the graph and extend the catchment basins. As a result one obtains a precise detection of the catchment basins and a graph of the steepest trajectories. A last iterative algorithm allows to follow selected downwards trajectories in order to detect particular structures such as rivers or thalweg lines of the topographic surface.

The watershed concept and its use in segmentation : a brief history

The watershed is one of the most used tools in image segmentation. We present how its concept is born and developed over time. Its implementation as an algorithm or a hardwired device evolved together with the technology which allowed it. We present also how it is used in practice, first together with markers, and later introduced in a multiscale framework, in order to produce not a unique partition but a complete hierarchy.