Compact Hash Codes and Data Structures for Visual Descriptors Retrieval

This thesis is structured as follows. We begin by introducing a brief survey of related work on feature hashing for retrieval using local and global visual features illustrating all related methods (Chapter 2). In Chapter 3 we describe the standard components of an image retrieval system, including common performance measures for evaluation purposes. In Chapter 4 we introduce and describe a new approach for vector quantization based on kmeans which allows the possibility of assignment of a visual feature to multiple cluster centers during the quantization process. We show the goodness of this new approach by presenting an exhaustive comparison with all the methods presented in Chapter 2. We also introduce the usage of an efficient and recursive data structure to store datas. Finally, in the Chapter 5, we introduce a new approach for efficient image retrieval based on the m-k-means hashing introduced in Chapter 4 and we apply our method on CNN features. To conclude we introduce the usage of an efficient indexing structure based on Bloom Fiters and we show how the experimental validation outperforms the state-of-the-art hashing methods in terms of precision.