Large-scale Content-based Visual Information Retrieval

Rather than restricting search to the use of metadata, content-based information retrieval methods attempt to index, search and browse digital objects by means of signatures or features describing their actual content. Such methods have been intensively studied in the multimedia community to allow managing the massive amount of raw multimedia documents created every day (e.g. video will account to 84% of U.S. internet traffic by 2018). Recent years have consequently witnessed a consistent growth of content-aware and multi-modal search engines deployed on massive multimedia data. Popular multimedia search applications such as Google images, Youtube, Shazam, Tineye or MusicID clearly demonstrated that the first generation of large-scale audio-visual search technologies is now mature enough to be deployed on real-world big data. All these successful applications did greatly benefit from 15 years of research on multimedia analysis and efficient content-based indexing techniques. Yet the maturity reached by the first generation of content-based search engines does not preclude an intensive research activity in the field. There is actually still a lot of hard problems to be solved before we can retrieve any information in images or sounds as easily as we do in text documents. Content-based search methods actually have to reach a finer understanding of the contents as well as a higher semantic level. This requires modeling the raw signals by more and more complex and numerous features, so that the algorithms for analyzing, indexing and searching such features have to evolve accordingly. This thesis describes several of my works related to large-scale content-based information retrieval. The different contributions are presented in a bottom-up fashion reflecting a typical three-tier software architecture of an end-to-end multimedia information retrieval system. The lowest layer is only concerned with managing, indexing and searching large sets of high-dimensional feature vectors, whatever their origin or role in the upper levels (visual or audio features, global or part-based descriptions, low or high semantic level, etc. ). The middle layer rather works at the document level and is in charge of analyzing, indexing and searching collections of documents. It typically extracts and embeds the low-level features, implements the querying mechanisms and post-processes the results returned by the lower layer. The upper layer works at the applicative level and is in charge of providing useful and interactive functionalities to the end-user. It typically implements the front-end of the search application, the crawler and the orchestration of the different indexing and search services.