Home | Research | Publications|
Teaching
| Students | Links
| Engineering | Bio
Patrice
Brault's
Homepage
L2S,
CentraleSupelec, Paris-Saclay University.
____________________________________
CNRS Engineer-Researcher,
L2S, Laboratory of Signals and Systems, CNRS UMR
8506
CentraleSupelec, Plateau de Moulon, 3 rue Joliot-Curie,
91192, Gif sur Yvette Cedex, FRANCE.
Room :
C4.23 Phone: +33 1
69 85 17 73
Fax :
E-mail: patrice.brault@l2s.centralesupelec.fr
Researcher ID :
GoogleScholar :
https://scholar.google.fr/citations?user=Ps34jaoAAAAJ&hl=fr
ResearchGate :
https://www.researchgate.net/profile/Patrice_Brault
IEEE Senior Member (SM'2009)
ORCID 0000-0002-3653-7307
-
- Main research interests and skills
: Image and Video Processing, Analysis and
Compression, Discrete and Redundant Wavelets (Motion-tuned, Coherent States), Motion Estimation, Segmentation, Semantic
Scene Analysis, Shape Recognition, Bayesian
Inference, Variational methods, Fractal Analysis, Physics of solids (Quantum Mechanics, Quantum
Electronics) ...
---------------------------------------------------------------------------
- Examples of DNNs (R-CNN and YOLO) Semantic and instantiated
segmentations initiated by an optical flow motion estimation
(Farneback Algorithm).
- Instantiated
segmentation by RCNN mask_Cars on Highway.mp4 (download
to view)
- Instantiated
segmentation by R-CNN mask.mp4 (download to view)
- Semantic segmentation initiated by Optical Flow.mp4
- Image of an instantiated semantic segmentation with representation
of the speed vectors (argument and module) by a color (after the
color-speed graphic representation below)
In 1, 2 and 3 the semantic value is given by
the name of the object : car, truck, pedestrian, bag, bicycle. The real
number indicates the degree of confidence in the semantic value.
In 1 and 2, the objects are instantiated by
a different color.
In 3, the color indicates the orientation
and the absolute value of the speed given by the Optical Flow.
This one initiates and speeds
up the segmentation which results in a joint Motion Estimation-Semantic
segmentation.
Color-speed graphic
representation
-------------------------------------------------------------------------------------
Master
2 Internship subjects proposals (
2022-2023)
1) Semantic instantiated segmentation initiated by a motion
estimation (ME).
This subject started beeing explored in 2019. The
semantic segmentation has been realized by means of learning
methods : first an RCNN architecture then the YOLO algorithm. This
last one enables a fast instantiated segmentation. In order to
accelerate these segmentations, we initiate them by a non-learning
method of motion estimation : the classical optical flow algorithm
(OF). The interest now is to implement a learned motion estimation
and integrate
it in the same deep neural network. Experiments
have been done on road trafic sequences and city pedestrians,
bicycles and cars ones (see above on this page).
Semantic_Segmentation_and_MotionEstim_with
DNNs_M2R Subject_2019-2020.pdf
2) Image Quality Assessment (IQA) :
Objective Contrast Measurement by learning (Deep Neural
Nets)
We, in a former approach, have
used a mathematical approximation by a 3 variables polynomial,
and an appropriate parametrization, of the MOS (Mean Objective
Score) curve. The goal was to be able to reproduce, with a good
confidence, the MOS given by a group of human subjects, by using
only the intrinsic parameters of the image (the variables of the
polynomial) that are : the Dynamic Range Occupation (DRO) of the
Luminance, the Histogram Shape Deformation (HSD) during contrast
enhancement and the Pixel Uniformity (PU) in a close
neighborhood inside the image.
In most
cases this model (Maximum Contrast Minimum Artifact, MCMA)
provides and objective result of the contrast quality evaluation
close to the MOS.
More
recently we turned to use Learning to reach the same, or better,
results. Interesting results have been obtained with
VGG16, a CNN architecture, and TID2013 and CEED2016 databases.
The BRISQUE evaluator has also been used. We observed that in
some cases the result is NOT in concordance with the results of
the MOS. In fact this may come from the perceptual property of
the human eye and Human Visual System (HVS). In fact it
seems that three other factors should be taken into account in
the computation of this task with a neural network, that are :
the spatial information (SI), the colorfulness (CF) and the
Global Contrast Factor (GCF)
Objective_Constrast_Measurement_M2R_subject_2019-20.pdf
3) Color restoration for underwater
images and videos
An imperfectly solved problem in underwater images is to
compensate the effect of depth on images, that tends to cancel all
colors and turn the major color to blue or green with depth. A
good solution has been to use an immerged multi color chart ( for
example the Mac Adam ellipse) to quantify the effect of depth and
colors filtering by the water. Then it is possible to make a quite
good correction on the images and videos.
We now would
like to test a learning algorithm to see if it is possible to do
the color correction without using this color chart.