Seamless remote browsing and coarse-to-fine compressed retrieval using a scalable image representation

In today's information society, two important trends can be observed : (1) the digital universe -particularly still or moving images- is growing exponentially and (2) computer networks are becoming more heterogeneous. The former raises the need for tools to guide the user within large data spaces, while the latter requires a representation of information that is sufficiently flexible to adapt itself optimally to various network constraints or user requirements. In this thesis, we focus on image information and address these needs by investigating three complementary aspects of a remote image server. First, a scalable image representation, namely JPEG2000, is studied. This compression standard organizes the image information in such a way that almost any spatial area can easily be extracted at any resolution and bit-depth. This scalability implies a high computational load that may require a hardware implementation when dealing with real-time constraints. The main characteristics of such hardware architecture are also presented. Secondly we consider the issues of scheduling and caching JPEG2000 data in client/server interactive browsing applications under memory and channel bandwidth constraints. We evaluate several strategies to schedule data packets that are likely to become relevant to expected future Windows-of-Interest ("pre-fetching" techniques), and show how the system reactivity can be improved. Finally, we investigate how to perform categorization and retrieval tasks directly in the compressed domain. Thanks to the JPEG 2000 scalability the discriminant level of a compressed image characterization is directly related to the amount of extracted data. Taking this finding into account, an original representation, called integral volumes, is introduced to store such characterization. Combining integral volumes with random decision trees, we propose a JPEG 2000 image classifier that achieves performances similar to the best uncompressed image classification results obtained on several freely available databases. Eventually, to address the image retrieval problem, a cascade of such classifiers is used, enabling a cost-optimized coarse-to-fine retrieval process.(FSA 3) -- UCL, 200