The OBELIX team from IRISA is organizing a Workshop on Hierarchical Image Representations, Computer Vision and Remote Sensing.
This event will be held in Vannes (Campus of Tohannic) on January 25th, 2016.
The workshop will consist in 4 invited talks (see details below) and will take place in the ENSIBS building, Main Amphitheater from 9:30 to 12:30. A welcome coffee will be available from 9:00 and a lunch will be offered under registration. For the sake of organization, registration is mandatory (before January, 18th): registration link
After lunch, Petra Bosilj will defend her PhD entitled “Image Indexing and Retrieval using Component-Trees”.
- General and Efficient Scene Text Localization and Recognition (Jiri Matas)
End-to-end text localization and recognition in the wild, a.k.a scene text detection, is an open problem with many applications. First, I’ll review formulations of the text detection problems and briefly overview the state of the art. In the rest of the talk, I will introduce a novel text detection and recognition method. Exploiting the observation that text in virtually any script is formed of strokes, the method starts by detecting text fragments, stroky regions that could correspond to either characters, their parts or their groups, such as whole words. Such generalization of character detection makes the method applicable to a wide variety of scripts including Latin, Hebrew, Arabic or Hindi, etc.) and fonts. The detection stages of the proposed pipeline are scale- and rotation- invariant and do not exploit a language model. After a standard line-formation step, recall is increased by a text-line resegmentation based on a modified graph-cut algorithm. The resegmentation allows to discard the standard assumption of region-based methods that all characters are detected as connected components. The method runs in real time and achieves state-of-the-art text localization and recognition results on the ICDAR 2013 and 2015 Robust Reading datasets.
Jiri Matas is a full professor at the Center for Machine Perception, Czech Technical University in Prague. He holds a PhD degree from the University of Surrey, UK (1995). He has published more than 200 papers in refereed journals and conferences. Google Scholar reports about 22 000 citations to his work and an h-index 54. He received the best paper prize at the International Conference on Document Analysis and Recognition in 2015, the Scandinavian Conference on Image Analysis 2013, Image and Vision Computing New Zealand Conference 2013, the Asian Conference on Computer Vision 2007, and at British Machine Vision Conferences in 2002 and 2005. His students received a number of awards, e.g. Best Student paper at ICDAR 2013, Google Fellowship 2013, and various “Best Thesis” prizes. J. Matas is on the editorial board of IJCV and was the Associate Editor-in-Chief of IEEE T. PAMI. He is a member of the ERC Computer Science and Informatics panel. He has served in various roles at major international conferences, e.g. ICCV, CVPR, ICPR, NIPS, ECCV, co-chairing ECCV 2004 and CVPR 2007. He is a program co-chair for ECCV 2016. His research interests include object recognition, text localization and recognition, image retrieval, tracking, sequential pattern recognition, invariant feature detection, and Hough Transform and RANSAC-type optimization. For more, see http://cmp.felk.cvut.cz/~matas.
- Binary Partition Tree for remote sensing applications (Philippe Salembier)
This talk will discuss the interest of Binary Partition Trees (BPT) for remote sensing applications. BPTs can be considered as an initial abstraction from the signal in which raw pixels are grouped by similarity to form regions, which are hierarchically structured by inclusion in a tree. They provide multiple resolutions of description and easy access to subsets of regions. Their construction is often based on an iterative region-merging algorithm. This approach and the associated notions will be discussed. Several examples of construction will be illustrated for synthetic aperture radar (SAR) and hyperspectral images. Once constructed, BPTs can be used for many applications including filtering, segmentation, classification and object detection. Many processing strategies consist in populating the tree with features of interest for the application and in applying a specific graph-cut called pruning. These ideas will be illustrated in particular for hyperspectral image classification and for SAR speckle noise filtering and segmentation.
Philippe Salembier received an engineering degree from the Ecole Polytechnique, Paris, France, in 1983 and an engineering degree from the Ecole Nationale Supérieure des Télécommunications, Paris, France, in 1985. From 1985 to 1989, he worked at Laboratoires d’Electronique Philips, Limeil-Brevannes, France, in the fields of digital communications and signal processing for HDTV. In 1989, he joined the Signal Processing Laboratory of the Swiss Federal Institute of Technology in Lausanne (EPFL), Switzerland, to work on image processing and received a PhD in 1991. At the end of 1991, after a stay at the Harvard Robotics Laboratory as Postdoctoral Fellow, he joined the Technical University of Catalonia, Barcelona, Spain, where he is currently professor lecturing in the area of digital signal and image processing. His research interests include image and video sequence processing, compression and indexing, mathematical morphology, level sets and nonlinear filtering, as well as remote sensing image processing and signal processing tools for genomics.
- A “Diplomatic” Component Tree Algorithm for Extreme Dynamic Range Images (Michael Wilkinson)
The standard parallel approach to building component trees uses a spatial division of the data into disjoint sections. It then builds a component tree and hierarchically merges the trees into a single tree. The merging stage is similar to that of merge sort, once for each pair of pixels on the boundary separating two disjoint regions. This means the time complexity rises linearly with the number of grey levels in the image, or exponentially with bit depth. Experiments have shown that up to about 16 bpp the algorithm provides good speed up. Beyond that, its speed-up rapidly drops below unity. To counter this problem, we propose a hybrid algorithm that avoids the expensive merging stage, by first computing a pilot component tree on a discretized version of the image using a root-to-leaf flooding approach, followed by a refinement phase, where the data are distributed by grey level rather than spatially. This algorithm provides a speedup of up to about 40 on 64 cores.
Michael Wilkinson obtained an MSc in astronomy from the Kapteyn Laboratory, University of Groningen in 1993, after which he worked on image analysis of intestinal bacteria at the Department of Medical Microbiology, University of Groningen, obtaining a PhD at the Institute of Mathematics and Computing Science, also in Groningen, in 1995. He was appointed as researcher at the Centre for High Performance Computing in Groningen working on simulating the intestinal microbial ecosystem on parallel computers. During that time he edited the book “Digital Image Analysis of Microbes” (John Wiley, UK, 1998) together with Frits Schut. After this he worked as a researcher at the Johann Bernoulli Institute for Mathematics and Computer Science (JBI) on image analysis of diatoms. He is currently senior lecturer at the JBI, working on morphological image analysis and especially connected morphology, with applications ranging from the quantum to the cosmological scales.
- Vector attribute profiles for hyperspectral image classification (Erchan Aptoula)
Morphological attribute profiles are among the most prominent spectral-spatial pixel description methods. They are efficient, effective and highly customizable multi-scale tools based on hierarchical representations of a scalar input image. Their application to multivariate images in general, and hyperspectral images in particular, has been so far conducted using the marginal strategy, i.e. by processing each image band (eventually obtained through a dimension reduction technique) independently. In this talk, we investigate the alternative vector strategy, which consists in processing the available image bands simultaneously. The vector strategy is based on a vector ordering relation that leads to the computation of a single max- and min-tree per hyperspectral dataset, from which attribute profiles can then be computed as usual. Known vector ordering relations for constructing such max-trees and subsequently vector attribute profiles are explored, and a combination of marginal and vector strategies is introduced. An experimental comparison of these approaches in the context of hyperspectral classification with common datasets is also provided, where the proposed approach outperforms the widely used inal strategy, i..