This doctoral research project is a joint project between SEASIDE and TEXMEX teams from IRISA laboratories. Prof. Sebastien Lefevre and Dr. Ewa Kijak have respective recognized expertise in morphological image analysis (from which component trees are issued) and in multimedia indexing (especially image indexing).
Component trees aim at representing a digital image by the regions it contains at various scale, through a tree-based structure. Their invariance properties (to translation and rotation) and their robustness to noise (Perret, 2010) have motivated some recent works in image indexing (Alajlan et al., 2008 ; Urbach et al., 2007 ; Vilaplana et al., 2008), but their usage in this field stays limited. The goal of this research project is study more deeply their use and their abilities in the context of content-based image indexing and retrieval (CBIR). Several ways are considered to do so.
The most straightforward approach would be to consider the component-tree of an image as its global descriptor, since the tree is a way to represent the image. The problem to solve is then to be able to compare global descriptors, which can be stated as a tree matching problem. Studies will focus on the efficiency of such a comparison or matching, and the possible design of some transforms in order to represent these tree-based structures as vectors, which are the usual representation in CBIR.
It is also possible to consider component-trees in the context of image indexing through local descriptors. Trees may be used here either to detect interest regions, or to describe them. Indeed, each branch of a component-tree represents an image region described at various scales (a node of the branch being related to an observation scale of the region). Since the tree of an image may be used to extract interest region at various scales, it appears as an original and relevant solution for interest region description.
Besides, each region of a tree can be described by some features (from simple luminance measures to more complex descriptors, which despite being more adapted to image retrieval, may require to adapt the tree structure). A component-tree can then be used for both region detection and description. Using such a representation tool is also of high interest since it is intrinsically scalable: each node/region is not only described by its own features but also by those of its descending nodes.
In order to ensure scalability, usage of component-trees may be further analyzed by studying how they can lead to a sparse representation, without the need of local description and aggregation steps. Sparse representations have recently shown their interest in image indexing (Sivic et al., 2009 ; Zepeda Salvatierra, 2010). Moreover, tree-based structures gather some connectivity information between regions. Such information may be used to avoid to rely on the usual post-processing geometric robustification step.
Experiments on image retrieval will rely on the multimedia indexing facilities from TEXMEX team within IRISA. An operational testing environment is available, which contains many image databases (with several millions of images) and which offered efficient retrieval algorithms. This will help to perform experimental evaluations in the context of image search (similar scenes, similar objects) or copy detection. Moreover, description and indexing algorithms might be implemented on the distributed architecture from either SEASIDE or TEXMEX team.