December 21st, 2011 @ 1130am Matthieu Cord (Sorbonne)

In this talk, I briefly review the research activities and projects of my multimedia group in the MALIRE/LIP6 (Machine Learning and Information Retrieval) lab. We are mainly contributing in the area of image understanding by designing complex representations, kernel similarities and learning hierarchical and bio-inspired architectures. We are developing systems both for image retrieval and for classification and detection. First, I present the main components of our content-based image and video retrieval system — extraction and comparison of features, search by similarity, relevance feedback — focusing on interactive-learning based approaches. Several issues concerning scalability, including fast kNN search methods to speed up interactive retrieval, are discussed. Second, I present an application on text detection and recognition in real images taken on unconstrained environments, which remains surprisingly challenging in Computer Vision. The SnooperText strategy, combining bottom-up and top-down mechanisms to detect text boxes, is introduced. The bottom-up part is based on character segmentation, classification and grouping . The top-down part is achieved with a statistical learning approach based on Fuzzy HOG box descriptors, fully adapted for text box analysis. Experiments and an application to a GIS system with a (fully) automatic textual indexing are presented.

seminars/seminaritems/2011-12-21.txt · Last modified: 2011/12/16 00:31 by silberman