| With the digital image/video production and distribution industry sectors continuing to grow, visual data are available everywhere in our daily lives. It is essential to develop techniques that enable users to easily access large volumes of data. In this dissertation, we present computationally efficient techniques to represent, match, and index various types of visual data, with the primary goal of enabling effective and efficient search by either examples or keywords. First, we develop a new representation based on an ordered list of feature descriptors. Such a globally ordered, and locally unordered, representation is more discriminative than a conventional bag-of-features model because the descriptors' order offers additional information. Secondly, we propose using approximate sequence matching for comparing similarities between such representations. This method solves a correspondence problem between two ordered sets of features and calculates the similarities between both the matched and the unmatched features. This matching approach can also be used for automatically searching for the best alignment between two pieces of visual data. Furthermore, we propose acceleration methods for the matching process without compromising accuracy. Extensive experiments on various visual retrieval and classification tasks demonstrate the superior performance of the proposed techniques compared to existing solutions. Finally, this dissertation includes a thorough experimental study to understand the efficacy of state-of-the-art recognition methods for image annotation in real-world scenarios. The findings of this research should provide information for various design choices that can be used to create a practical, highly accurate retrieval system. |
