Interactive, incremental and multi-level exploration of large collections of images. ( Exploration interactive, incrémentale et multi-niveau de larges collections d'images)

The research work that is presented and discussed in this thesis focuses on large and evergrowing image collections. More specifically, we aim at providing one the possibility to explore such image collections, either to extract some kind of information and knowledge, or to wander in the collections. This thesis addresses this issue from the perspective of Interactive Data Exploration and Analytics. We take advantage of the similarity-based image collection browsing paradigm and aim at meeting simultaneously the three following constraints: (i) handling large image collections, up to millions of images, (ii) handling dynamic image collections, to deal with ever-growing image collections, and (iii) providing interactive means to explore image collections. To do so, we jointly study the indexing and the interactive visualisation of large and ever-growing image collections. Our contribution is three-fold. First, we focus on the incremental construction of a proximity graph, namely the Relative Neighbourhood Graph (RNG), to structure the image collections. Second, we propose a hierarchical and graph-based hybrid viewable structure to allow an interactive exploration of large image collections. A data partitioning algorithm, namely BIRCH, is used to yield this structure. Last, we present our interactive visualisation platform for large image collections. To evaluate the platform, along with an user evaluation, several different use cases are discussed such as a recommendation system for ever-growing image collections and visual analysis system for Document Image Analysis. These use cases are based on real world large image collections, that contain up to seven millions of images.