Player is loading...

Embed

Copy embed code

Transcriptions

Note: this content has been automatically generated.
00:00:01
ah hello everyone
00:00:03
i'd like to present my group of other computational by you in e. g.
00:00:10
so here's what i'm gonna talk about or did you were you were you what
00:00:13
were doing uh i'll be talking first about what actually computationally meeting is
00:00:18
you know i'll talk a bit about how we do my cross you the
00:00:21
ball developing hearts and on it should use e. g. apps imaging platform
00:00:27
and then i'll give you an idea about how we can get sharper images uh with the use of a artificial intelligence
00:00:34
here's my group ah actually there's a two members wrap posters outside so encourage you to over
00:00:41
lunch break to go in and see them olivia and uh if you and not
00:00:46
h. zero posters in uh in the second so what is computational imaging our
00:00:50
goal is actually to make images based on data that we collect
00:00:56
and so the definition here that's used by recently watched journal
00:00:59
is uh that it uses innovative algorithms computation sensor design
00:01:05
to dress image formation problems so it's very close to
00:01:08
image processing and actually to compared to uh the
00:01:11
inputs in the outputs might similar uh image processing typically you take an image given other image out
00:01:18
in our case we have also uh collect data or it could be images
00:01:24
and also get an image out we try and also incorporate some knowledge about how
00:01:29
this image has been collected in order to get the better image yeah
00:01:33
concrete example image processing you can take a bad image as an input and get a much better image
00:01:40
as an output by using your photoshop insert grammar or whatever your favourite image processing to liz
00:01:46
ah in our case uh we're we're looking into trying to figure
00:01:50
out things in seeing things that wouldn't be directly accessible
00:01:54
with your typical imaging device so in our case that could
00:01:58
be tomorrow if he were you start out with projections
00:02:02
so you look sense you could be similar as with x. rays you could look
00:02:06
through a a body but you cannot actually slides however you want to
00:02:11
do is you can only do after you've actually done some computations
00:02:16
uh inverted uh the the image formation process in once you've done
00:02:21
so you can essentially look at your data for in
00:02:24
you can look at your data in all directions uh you can move it around and slice it however you want
00:02:32
so let me let me give you a one no uh one example of a project that that uh
00:02:38
we're working on and that's a collaboration that we have with the uh three groups were in france
00:02:44
use but to have a good critic need who's uh who's developing optical microscopes uh you
00:02:51
don't go more who's in a stress ball was the developmental biologist studying hard development
00:02:58
and so is a person an agenda that there was a she's in uh which isn't universe you burn
00:03:06
and so the goal there is to try and figure out how
00:03:08
the heart develops and uh why sometimes the development goes along
00:03:15
and one of the things that i'd like to be able to do is to actually be able to see
00:03:20
what happens in particular particular see uh what happens when things go wrong
00:03:25
what special with the heart it's a little bit different from other organs inner body is that
00:03:30
the hearts actually already functional which means that the heart is beating
00:03:34
long before it's fully formed and so it starts out
00:03:37
in the human looks a bit like it to be first and that tubes already being it's already pumping but
00:03:43
and this is typically when you go from here to here this is where a lot
00:03:47
of the effects might actually happen and it's very hard to image something that's
00:03:53
small and be that uh uh moving or beating very
00:03:58
quickly uh several uh heart beats per sec
00:04:02
so can we image so when you put and and in this case we're we're
00:04:06
looking at the uh an animal mall that's a zebra fish which is great
00:04:10
transparent and you can see all the organs without
00:04:13
doing anything special using any uh special
00:04:18
uh_huh nation just plain like
00:04:21
and when you look at a time lapse we can essentially folder whole development from the signals from a single cell all the way to
00:04:28
when the organ start forming here on top you can actually see by uh that's right here this
00:04:34
is the brain uh here you have the heart that's already bleeding and you can see that
00:04:38
can barely see although you can see the other organs forming in the
00:04:41
heart starts beating it's way too fast captured by a simple timeout
00:04:48
so that problem has pretty much everything that somebody who does
00:04:53
image processing or computational imaging uh likes to tap
00:04:57
we're looking at things that are very small uh some aspects are
00:05:01
slow development some aspects are fast the heart that's beating
00:05:08
original looking at objects like to get differentials shape you have things that you'd like to imagine multiple
00:05:14
channels the signal is very weak in noisy and the sample itself is very bad child
00:05:20
and of course if you want to do any valid here is this sixteen people to do experiments multiple time
00:05:27
this is pretty much the world that we're living in and we're trying to find the solutions solutions are trying to find
00:05:32
have to essentially a depending on on the day and the problem have to uh have to comply with this constraint
00:05:40
so let me that show you one of the problems that the actually has been working on a
00:05:46
uh recently and that's uh the idea of trying to make a small microscope that has some
00:05:51
very nice features namely is mike of microscopes out there that allow you to actually image
00:05:59
in three dimensions so get data from one points inside uh the inside
00:06:04
a a three dimensional object at a time that second focal microscope
00:06:09
something has been out there for a very long time which is great but
00:06:13
it has the major drawback that it's very slow because we get
00:06:18
one point of time so we have two centuries stand uh through your object in order to form an entire image
00:06:25
so now imagine if you're a teacher imaging the heart this is what actually
00:06:29
happens you have the heart here on the left that's leading and what
00:06:32
you're doing you're standing well you don't really get something that looks like a
00:06:35
heart get more something that um looks well it looks like this
00:06:40
so now the question was can we take those images that are horribly a matter rated
00:06:46
and try and reconstruct something that looks more like a true shape of the heart so
00:06:51
this is what uh the kind of images that only yes starts with it
00:06:56
and you can see that the heart is is uh where there's something happening those
00:06:59
images but uh it doesn't look anything like a part two should look like
00:07:08
and then finally uh when you when we do the reconstruction uh were able to uh
00:07:13
which is able to actually put the images back into the right word or
00:07:17
and to onto this that's standing artifacts to actually get the heart of beat as it should be beating
00:07:27
another problem that we're facing a so i'm i mention all the challenges that we have
00:07:32
one thing that's expected from uh the imaging community is that uh we
00:07:38
other people are able to actually produce the results with the the idea is that we haven't publish
00:07:43
and uh there's some alarming don't use that and many of those
00:07:47
results a general cannot be reproduced at least in the
00:07:51
biomedical science i think that in majoring in computer science uh this is something that's very easy
00:07:57
to do because we you know we are browns and the about is what turns out
00:08:03
it's actually not always all that easy especially because uh
00:08:07
often very often the data that you collect
00:08:10
uh has multiple purposes a very often uh your your uh
00:08:16
you know it's very specific they that that you get from a collaborator who's working biology
00:08:21
who has a different agenda about when and where you'd like to should like to to publish that data
00:08:27
and so uh the other thing is that if you actually want to
00:08:32
distribute these large data set it actually uh requires lots of
00:08:35
resources because you don't just need to do it for six months but you have to do it on the on the long
00:08:42
so we've uh uh we we've got two colleagues at any gaps so it's it's much general
00:08:47
that's the best enough that we decide to uh put together a platform and that platforms in imaging platform
00:08:55
that that looks like this that's down here in the basement
00:08:58
and that allows us to acquire dynamic that imaging data
00:09:03
with a robotic arms there is a more drives microscope that's it down here to
00:09:08
the microscope that's has been partly built here in valet with the help of
00:09:14
the mechanical uh apprentices and install and that allows us to uh collect data in a
00:09:21
very controlled manner and uh in the repeatable manner so that we can uh
00:09:27
for instance in this case we're we're looking at an algorithm for super resolution can image essentially the same
00:09:32
motion multiple times with different kinds of cameras different
00:09:37
kinds of ruminations strategies and should we compare
00:09:40
a a tree compare the and two and a quality of our
00:09:44
algorithms uh to see how we're uh how we're too
00:09:48
and so uh uh show you the two arms that are there that that the reviews for that experiments
00:09:54
the one of the nice things that you country constrain them in in ways that the motion will um
00:10:01
will fall essentially either something that has been preprogrammed were that will explore certain space
00:10:06
within a certain number of of uh of uh of constraints that you set for
00:10:13
finally uh i'd like to talk about the problem that uh adrian has has been working on
00:10:19
and that's the idea uh another classical problem in computational imaging
00:10:24
is that of trying to people are blurry images
00:10:27
and uh when i said for what we do is collect images that
00:10:32
look like this one here on the right which is completely blurred
00:10:36
and if you're lucky enough to know what kind of microscope you've been
00:10:41
using and what's the four characteristic of that microscopes or microscope or
00:10:47
and that's typically uh inside this magical function that's yours called the point spread function
00:10:53
which pretty much cracked rises your entire optical system then there is hope that you might
00:10:58
eventually be able to pull back from this average image back to know original
00:11:04
so still that even if you have that functions a very difficult problem and now if you don't have that much but
00:11:10
function or if there's not one but many depending on where
00:11:14
you are in your image becomes an even harder pop
00:11:18
so uh what we in uh ask their cells is whether we could essentially
00:11:24
instead of having to meet your happens but function which is something that's very painful to do whether
00:11:28
we could have a program actually recognise when you give it an image recognise what kind of
00:11:36
a degradation was the source what what was the what was that points but
00:11:41
so here is how we tackle that problem we've taken a very big library of good images
00:11:48
and then uh we've parameter tries the point spread functions with decorations
00:11:53
and we've generated a a hole to have a lot of as possible decorations
00:11:58
and we've taken all these images and we've degraded them with different kinds of of course but
00:12:04
then we had this big library and uh we trained uh uh
00:12:09
i would betray a convolutional neural net work to actually be able when
00:12:13
you get it in in each without telling it's the parameters
00:12:18
that it will tell you what kind of parameters were actually uh what kind of parameters was a great were the parameters of the post
00:12:26
what you've trained that the nice thing is that you know can now take and on known image
00:12:32
that can have that in uh that was degraded in different places with
00:12:36
different ways but functions this case it's something that's flat that's inclined
00:12:41
respect to the microscope so some regions are in focus some other regions are not in focus
00:12:47
and you can send small patches into that uh into that and then that will don't
00:12:53
tell you oh in this region image has been degree of this kind of stuff
00:12:58
and then uh you can run a space apparently convolution can get an
00:13:02
image that now is word focus in all regions of your age
00:13:07
when you want a little bit more details about how are we doing this and how it performs 'cause visit uh engines post
00:13:14
just outside
00:13:16
so in summary ah computational imaging there's many applications
00:13:21
ranging now from entertainment to science were all have them in our cell phones a
00:13:26
group mostly focusing on applications uh in in the area of of science
00:13:31
for austin sounds to which usually that we're able to reproduce
00:13:34
things and crack tries things a very carefully otherwise
00:13:38
uh maybe we uh have an issue with our partners will will show us
00:13:44
people just like to see pretty images but the actual wanna know exactly whether
00:13:49
they also are meaningful
00:13:51
ah specifically in the area by you imaging uh been able to make slow
00:13:56
microscopes faster or teach uh did you computer to recogniser cracked arise
00:14:02
uh optical system which is something that's really painful if use you know lab never major points but function it's not
00:14:08
something that's actually on a to do and having a computer do that for you is a is a big
00:14:14
our imaging platforms actually uh it's open uh and uh no
00:14:21
welcomes actually poppers if you're interested in using that well
00:14:25
so would i like to acknowledge the you know the people actually found our our research the national science
00:14:31
foundation and also uh the value value some still initiative which has been a a launched by page
00:14:36
up a few years ago and it allowed us to allow me to recruit you know excellent students
00:14:41
who had been a train their their their elementary school in a in a in that it
00:14:48
and uh
00:14:50
bring them back to uh nothing and find the
00:14:55
any questions on show will help us build the uh

Share this talk: 


Conference Program

Introduction by Hervé Bourlard
BOURLARD, Hervé, Idiap Director, EPFL Full Professor
Aug. 29, 2018 · 9:03 a.m.
916 views
Presentation of the «Speech & Audio Processing» research group
MAGIMAI DOSS, Mathew, Idiap Senior Researcher
Aug. 29, 2018 · 9:22 a.m.
16949 views
Presentation of the «Robot Learning & Interaction» research group
CALINON, Sylvain, Idiap Senior Researcher
Aug. 29, 2018 · 9:43 a.m.
9721 views
Presentation of the «Machine Learning» research group
FLEURET, François, Idiap Senior Researcher, EPFL Maître d'enseignement et de recherche
Aug. 29, 2018 · 10:04 a.m.
14111 views
Presentation of the «Uncertainty Quantification and Optimal Design» research group
GINSBOURGER, David, Idiap Senior Researcher, Bern Titular Professor
Aug. 29, 2018 · 11:05 a.m.
3210 views
Presentation of the «Perception and Activity Understanding» research group
ODOBEZ, Jean-Marc, Idiap Senior Researcher, EPFL Maître d'enseignement et de recherche
Aug. 29, 2018 · 11:24 a.m.
5620 views
Presentation of the «Computational Bioimaging» research group
LIEBLING, Michael, Idiap Senior Researcher, UC Santa Barbara Adjunct Professor
Aug. 29, 2018 · 11:45 a.m.
4130 views
Presentation of the «Natural Language Understanding» research group
HENDERSON, James, Idiap Senior Researcher
Aug. 29, 2018 · 2:03 p.m.
8976 views
Presentation of the «Biometrics Security and Privacy» research group
MARCEL, Sébastien, Idiap Senior Researcher
Aug. 29, 2018 · 2:19 p.m.
6512 views
Presentation of the «Biosignal Processing» research group
RABELLO DOS ANJOS, André, Idiap Researcher
Aug. 29, 2018 · 2:43 p.m.
4027 views
Presentation of the «Social Computing» research group
GATICA-PEREZ, Daniel, Idiap Senior Researcher, EPFL Adjunct Professor
Aug. 29, 2018 · 2:59 p.m.
7158 views

Recommended talks

Methods for Rule and Knowledge Extraction from Deep Neural Networks
Keynote speech: Prof. Pena Carlos Andrés, HEIG-VD
May 3, 2019 · 9:10 a.m.
1443 views