Interests

Scalability issues on Machine Learning and Classification

Currently, I am working on several topics related to scalabity issues on indexing, classification, machine learning, indexing and knowledge discovery. I am particularly interested on tasks like

• kNN search on Giga-records databases;
• non-supervised learning (clustering) on large (>10⁶ records), high-dimensional (>100 dimensions) databases;
• multimodal information retrieval over millions of documents.

High-dimensional indexing

The subject of my thesis and of my recent research is the indexing of high-dimensional multimedia descriptors in order to accelerate the k nearest neighbours search (also known as kNN search or similarity search). I am studying how the use of multiple moderate-dimensional indexes can help to tame the "curse of dimensionality" and how the use of space-filling curves can help to store the indexes into fast, convenient and easy-to-update data-structures like the one-dimensional B-tree.

Related publications:

• Multicurves, a method for high-dimensional indexing based on multiple moderate-dimensional space-filling curves is explained on the paper “High-Dimensional Descriptor Indexing for Large Multimedia Databases”, published on CIKM'08.
• Projection KD-Forests, a precursor based on multiple moderate-dimensional KD-trees is explained on the paper “Fast Identification of Visual Documents Using Local Descriptors”, published on DOCENG'08.
• The 3-Way Trees, a method based on redundant overlapping covering in a single KD-tree is explained on the paper “3-way Trees: A Similarity Search Method for High-Dimensional Descriptor Matching”, published on ICIP'07.
• Of course, all three methods are explained in detail in my Ph.D. thesis “Local-descriptor matching for image identification systems”.

Content-based information retrieval (CBIR) and Image Identification

My thesis was also concerned on the application of CBIR to image identification (also known as copy detection, near-duplicate detection — terms which I avoid, because the target images, in my case, may have suffered strong modifications. This technique finds application in the detection of copyright violations and also — which was my main interest — in many Cultural Heritage activities.

Related publications:

• A complete, but concise, explanation of the problem can be found in our paper “Fast Identification of Visual Documents Using Local Descriptors”, published on DOCENG'08.
• Our first experiments on the subject of image identification were reported in the paper “Content-based Retrieval of Images for Cultural Institutions using Local Descriptors” published on GMAI'06.
• There is a report on our participation in the French competition ImagEVAL, still at an early stage of our research on multimedia indexing.
• Again, my Ph.D. thesis “Local-descriptor matching for image identification systems”, discusses the matter to some detail, especially in the chapters 1 and 2.

Computer Sciences and Cultural Heritage

I am very fond of research in the interface between Computer Science and Cultural Heritage. My M.Sc. research was related to the serious (and still unanswered) problem of Digital Longevity: how to preserve digital data for decades and even centuries. During my Ph.D. my emphasis shifted from preservation to access: the use of CBIR to allow the retrieval of images whose metadata are missing. I am still interested in all problems pertaining to this rich interface: digital preservation, digitisation of collections, digital libraries, asset management for Cultural Heritage, and digital techniques for conservation / restoration.

Related publications:

• A description of my Ph.D. work addressed to the Cultural Heritage community can be found in “Content-based Image Identification on Cultural Databases”, published on ICHIM'07.
• My pre-Ph.D. work, which concerned digitial longevity and digitization, is almost exclusively reported in Portuguese. The most significant of those reports is my M.Sc. dissertations, about the use of multimedia information systems on permanent collections, with an extensive chapter (chapter 3) on the digitization of photography and other 2D material, and another extensive chapter (chapter 5) on the longevity of digital media.