ISP Seminars

Previous Seminars

On the use of 4D imaging in cancer treatment
Guillaume Janssens
24 October 2011
This presentation will give a description of the multiple steps and image processing tools needed in the complex workflow that makes use of mutli-modal 4D imaging techniques to provide the needed insight for radiotherapy treatment of cancer.

-- Monday, 24 October 2011 at 10:45 (45 min.)

The use of sparsity hypothesis for source separation
Prasad Sudhakar
13 October 2011
Sparsity has been traditionally exploited for data compression, and a popular hypothesis to solve under-determined linear inverse systems (sparse recovery problem). Considerable amount of work has been done, both theoretical and algorithmic, in this regard. Of late, sparsity is being used to perform more complicated tasks such as source separation, learning, etc. The focus of this talk will be the usage of sparsity hypothesis for source separation.

-- Thursday, 13 October 2011 at 10:45 (45 min.)

Gait feature extraction in Parkinson's disease using low-cost accelerometers
Julien Stamatakis
17 June 2011
There are currently no validated clinical instruments or device that allow a full characterization of gait disturbances in Parkinson's disease. As a step towards this goal, a four accelerometer-based system is proposed to increase the number of parameters that can be extracted to characterize parkinsonian gait disturbances such as Freezing of Gait or gait asymmetries. After developing the hardware, an algorithm has been developed, that automatically epoched the signals on a stride-by-stride basis and quantified, among others, the gait velocity, the stride time, the stance and swing phases, the single and double support phases or the maximum acceleration at toe-off, as validated by visual inspection of video recordings. The results obtained in a PD patient and a healthy volunteer will be presented.

-- Friday, 17 June 2011 at 14:30 (45 min.)

A very short introduction to digital topology
Sébastien Lugan
1 June 2011
After a quick definition of basic key concepts (adjacency, connexity, Jordan's curve theorem, etc.) of digital topology, we will quickly explore some main features and applications of digital topology for 2D/3D image processing and some of its possible extensions.

-- Wednesday, 1 June 2011 at 10:30 (45 min.)

Imaging techniques for LOW-DOSE and COLOR tomography: new challenges for Cone-Beam CT and PET with hybrid pixels
Prof Yannick Boursier (Université Aix-Marseille 2 / ESIL)
18 May 2011
Cone Beam Computerized Tomography (CBCT) and Positron Emission Tomography (PET) Scans are medical imaging devices that require solving ill-posed inverse problems. The models considered come directly from the physics of the acquisition devices, and take into account the specificity of the Poisson noise. We propose various fast numerical schemes to compute the solution. In particular, we show that a new algorithm recently introduced by A. Chambolle and T. Pock is well suited in the PET case when considering non differentiable regularizations such as total variation or wavelet l1-regularization. Numerical experiments indicate that the proposed algorithms compare favorably with respect to well-established methods in tomography. Moreover, the new generation of hybrid pixel detectors working in a photon counting mode makes possible to select photons according to their energy, thus offering the opportunity to go beyond the classical reconstruction techniques that provides an absorption reconstruction volume. We will study the open questions of reconstructing colored X-ray volumes.
Ressources :

-- Wednesday, 18 May 2011 at 10:30 (45 min.)

Robust point correspondence based on template matching for image registration
Kaori Hagihara
4 May 2011
I introduce two algorithms presenting robust methods for automatically matching points over two images for image registration. The subject scene is considered to be more or less planar or in distance so that the image transformation can be roughly approximated by a homography. The basic principle is local correlation measurement by template matching, however many incorrect matches or outliers remain with real images. In order to resolve this problem, hierarchical schemes are proposed. One progressively evaluates all potential matches starting with non-local constraints, such as spatial consistency, global smoothness, and epipolar constraint, that should be approximately satisfied across the image. The other progressively estimates image distortions starting with translation, rotation, affine, and homography by random voting followed by variable template matching compatible with the estimated distortions.

-- Wednesday, 4 May 2011 at 10:30 (45 min.)

Recognition of sport players' numbers using fast color segmentation
Cédric Verleysen
20 April 2011
For a few years, automatic human identification has become a hot scientific topic in the field of computer vision. Recognition of players on a sport ground is a special case of human recognition that a computer needs to be able to understand how sport games proceed. More precisely, distinguishing between players' numbers is relevant to interpret autonomously a sport game using various tasks such as tracking of players, control of active cameras, event recognition, etc. The purpose of this presentation is to introduce some methods used to label automatically each player based on the recognition of the digit printed on the jersey. The first part of the seminar will explain how color segmentation can be applied to isolate numbers from the jersey. Afterwards, digit recognition will be focused. Our contribution to color segmentation uses a training phase during which colors on the jersey are learnt to improve the speed of the next segmentations. Knowing the principal colors of the jersey, color masks are successively applied which enables real-time isolation of the numbers, as opposed to most of the other current segmentation methods. Once the number is extracted, the decision about its value is taken, thanks to a feature-based classification. More precisely, discriminant features such as the ratio between the width and the height of the number, number of holes, projective histograms and some central moments are used to separate numbers among them. This combination between a short time online training that learns the colors of the jerseys and the use of an offline trained features-based classifier on the digits speeds up the recognition of sport players' numbers compared to existing other methods. Finally, the real-time efficiency of the proposed method is validated on a basketball match in which good rejection ratio is more than 95% and number recognition performance exceeds 70%.
Ressources :

-- Wednesday, 20 April 2011 at 10:00 (45 min.)

Ball detection and tracking in multi-view setting
Pascaline Parisot
30 March 2011
In the context of team sport events monitoring, the various phases of the game must be delimited and interpreted. In the case of a basketball game, the detection and the tracking of the ball are mandatory. First some common visual characteristics of a ball are presented. Then a ball detection framework based on the extracted knowledge on all the views are detailed. Finally, a graph-based tracking is explained. It consists in a temporal analysis of ball candidates and more specially in detecting ballistic trajectories among them.

-- Wednesday, 30 March 2011 at 10:30 (45 min.)

Scale & Affine Invariant Interest Point Detectors
Rémy Labbé
16 March 2011
Local features have been shown to be well suited to matching and recognition as well as to many other applications as they are robust to occlusion, background clutter and other content changes. The difficulty is to obtain invariance to viewing conditions. This SPS presents a scale and affine invariant detection algorithm test bench for particle disease localization and extraction in the context of infected hip prosthesis. Infections are observed on 2D X-ray images. The detector is based on the following results : (1) Interest points extracted with the Harris detector can be adapted to affine transformations and give repeatable results (geometrically stable). (2) The characteristic scale of a local structure is indicated by a local extremum over scale of normalized derivatives (the Laplacian). (3) The affine shape of a point neighborhood is estimated based on the second moment matrix.
Ressources :

-- Wednesday, 16 March 2011 at 10:30 (45 min.)

Solving imaging problem with Graph Cuts
Jérôme Plumat
4 March 2011
In this seminar, Jérôme will review the Graph Cuts theory and different Graph Cuts solutions for common image processing problems like image segmentation, volumetric data reconstruction in X-Ray parallel beam computed tomography, and image registrations.
Ressources :

-- Friday, 4 March 2011 at 14:30 (45 min.)