ISP Seminars

Previous Seminars

On the recent progress of object detection using neural networks
Maxime Istasse
18 April 2018
Since the success of convolutional neural networks for image classification on the ImageNet dataset in 2012, the computer vision community has made use of neural networks to raise the state-of-the-art performance for a wide range of problems, e.g. classification, segmentation, ROI detection, object detection, image captioning, ... In this seminar, after a brief introduction to the neural network paradigm, we will dicuss the recent progress in object detection through some particular models. We will observe that modern neural networks can learn to infer jointly the bounding box and the class of each object of a scene in an end-to-end manner.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 18 April 2018 at 15:00 (45 min.)

Hardware architecture for machine learning and image processing
Martin Lefebvre (ECS Group, ICTEAM, UCL)
29 March 2018
For some decades, efficient machine learning and image processing algorithms have been successfully developed to achieve a broad range of tasks, e.g. prediction, classification, compression, ... However, the hardware used for running those algorithms used to be inefficient, both in terms of computational performance and power consumption. In this seminar, we focus on the state-of-the-art implementation of electronic circuits making use of such algorithms. In particular, computational CMOS image sensors and deep- and machine learning processors are discussed. On the one hand, imagers focus on reducing the amount of data produced by the sensor either by performing in-pixel or embedded processing of the image, based on techniques such as edge detection, feature extraction, or compressive sensing. On the other hand, some cutting-edge deep- and machine learning processors rely on analog computations performed inside SRAM memories, to reduce the power consumption related to memory accesses and to increase the performance through parallelism, thus improving the energy efficiency of those processors.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 29 March 2018 at 16:15 (45 min.)

High-Performance Wireless Sensing with Low-Complexity Array Measurements
Dr.-Ing. Manuel S. Stein(Mathematics Department, Vrije Universiteit Brussel, Belgium,& Chair for Stochastics, Universität Bayreuth, Germany)
13 March 2018
While the technological capabilities regarding digital data transmission, storage, and computation have exponentially increased during the last decades, the advances associated with analog wireless equipment were only moderate. Therefore, today hardware cost and power consumption of radio interfaces form the main obstacles for constructing future wireless systems featuring either ultra-low complexity or ultra-high performance. However, in the advent of the Internet of things (IoT), where cheap and small devices are supposed to perform wireless sensing, and with the increasing demands for performance in critical infrastructure applications, it is inevitable further pushing radio technology towards these extremes.

Shannon Seminar Room (a105) Place du Levant 3, Maxwell Building, 1st floor -- Tuesday, 13 March 2018 at 11:00 (45 min.)

Tone mapping methods for X-ray images
Tahani Madmad
28 February 2018
Images are captured on a high dynamic range (HDR), divided into thousands of levels (typically 14 or 16 bits), when the sensor’s sensitivity allows it and when the application requires it. This is typically the case for X-ray radiography (medical or industrial), and this is a source of concern in terms of image rendering, since the human visual system perceives little more than 256 shades of gray.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 28 February 2018 at 14:00 (45 min.)

Greedy algorithms for multi-channel sparse recovery
Jean-François Determe (ULB & UCL, Belgium)
10 January 2018
Research on “compressive sensing” (CS) has shown that it is possible to retrieve high-dimensional signals using a limited set of (often random) linear measurements. For CS to work properly, the signal to be retrieved must be sparse, which means that most of its components (e.g., 90 % of them) are zero. Many algorithms relying on CS can reliably recover sparse signals on the basis of a number of measurements that essentially scales with sparsity level (instead of scaling with the number of dimensions).
Ressources :

Shannon Seminar Room (a105) Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 10 January 2018 at 11:00 (45 min.)

Computational Imaging in Atomic Force Microscopy
Thomas Arildsen (TPS/DES, Aalborg U., Denmark)
30 November 2017
Atomic force microscopy (AFM) is an imaging technique which can measure the surface structure of a specimen of interest down to nano-scale. It does this by scanning a tiny probe across the surface and thereby measuring a “height map” or other properties of the surface.

Shannon Seminar Room (a105) Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 30 November 2017 at 11:15 (45 min.)

Compressive learning (e.g., clustering) from a (quantized) sketch of the dataset
Vincent Schellekens
15 November 2017
Machine learning algorithms, such as the k-means clustering, typically require several passes on a dataset of learning examples (e.g., signals, images, data volumes). These must thus be all acquired, stored in memory, and read multiple times, which becomes prohibitive when the number of examples becomes very large. On the other hand, the machine learning model learned from this data (e.g. the centroids in k-means clustering) is usually simple and contains relatively few information compared to the size of the dataset.
Ressources :

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 15 November 2017 at 11:00 (45 min.)

Light Field Methods for the Visual Inspection of Transparent Objects
(invited talk) Johannes Meyer (KIT, Karlsruhe, Germany)
24 October 2017
Components made of transparent materials play an important role in human’s every-day life. For example, they are employed to build windshields for automobiles and aircrafts or to guide light beams in high-precision optical systems and therefore have to meet high quality requirements. However, powerful automated visual inspection systems for transparent materials are still an open research question.
Ressources :

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Tuesday, 24 October 2017 at 14:00 (45 min.)

Dedicated image processing tools for proton imaging
Sylvain Deffet
4 October 2017
Thanks to the finite range of the protons, proton-therapy is of undeniable therapeutic interest. However, proton treatment planning suffers from uncertainties in the computation of the protons range. As a means of measuring the proton energy loss along the path of the beam, proton radiography could be used to significantly increase the accuracy of proton therapy. We developed with this aim in mind a new prototype of proton imaging system based on multi-layer ionization chambers. The first tests on patients are planned end of this year.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 4 October 2017 at 10:00 (45 min.)

Deep Learning in Medical Imaging
Eliott Brion
29 March 2017
To illustrate its use in medical imaging, the goal of this seminar is to show how deep learning can automatically contour healthy organs and tumors in CT scans.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 29 March 2017 at 16:15 (45 min.)