PULSE: Photo Upsampling Makes Blurry Faces 60 Times Sharper

PULSE is a new image upsampling algorithm. This new algorithm transforms a blurry image into a high-resolution image.

It can take a super low-resolution 16x16 image and turn it into a 1080p high-definition human face! You don’t believe me? Then you can try it yourself in less than a minute using their demo.

Watch the video

References

Paper:https://lnkd.in/gUdHFJs​
Demo: https://lnkd.in/gmN3jTq​
Github:https://lnkd.in/gp5paYb

Video Transcript

00:00

this new algorithm transforms a blurry

00:02

image to a high-resolution image it can

00:04

take a super low resolution like a 16 by

00:07

16 image and turn it into a 1 ADP High

00:10

Definition human face you don't believe

00:12

me then you can do just like me and try

00:14

it yourself in less than a minute but

00:17

first let's see how they did that this

00:25

is what they I and I share artificial

00:27

intelligence news every week if you are

00:29

new to the channel and want to stay

00:31

up-to-date please consider subscribing

00:33

to not miss any further news to start we

00:36

have to introduce the concept of photo

00:38

up sampling our image super resolution

00:40

the goal here is to construct a high

00:43

resolution image from a correspondingly

00:45

low resolution input which is a face in

00:48

this case the low resolution will be

00:49

such as 16 by 16 pixels super blurry to

00:53

a high definition on one a DP image with

00:56

a clear face usually these techniques

00:58

use supervised learning to trained our

01:00

network and measure the average distance

01:02

between the new high-definition image

01:04

and a high resolution ground truth but

01:07

using this technique seem to neglect

01:08

important details like textures and

01:11

create some blurry spots when the answer

01:13

is uncertain because of the differences

01:15

between the low and high resolutions

01:17

this is where polls comes and play their

01:20

goal was to generate realistic images

01:22

within the set of plausible solutions

01:25

meaning that they wanted to rely on an

01:27

actual image that was realistic where

01:30

it's downscale version will look the

01:32

same as the original low resolution

01:34

image instead of having to guess

01:36

directly from the low resolution image

01:38

so they introduced a new self supervised

01:41

technique that traverses the high

01:43

resolution natural image manifold

01:45

searching for images that don't scale to

01:48

the original low resolution image this

01:51

is great because as you can see in this

01:53

image multiple high resolution images

01:55

can be done scaled to the same low

01:58

resolution image which will cause a

02:00

blurry uncertainty with previous

02:02

techniques these high resolution images

02:04

are made by a gun Network which is pre

02:07

trained in an unsupervised way to

02:09

generate multiple realistic and sharp

02:11

face images

02:12

if you are not familiar with self

02:14

supervised learning or generative

02:16

adversarial networks or Ganz I invite

02:18

you to watch the video I introduced them

02:20

both are linked in the description below

02:22

at first I was skeptical probably just

02:25

like you so I decided to try it myself

02:28

since the code is available and the

02:30

results were amazing you can set the

02:33

image is super blurry no one could guess

02:35

how I really looked like in this picture

02:37

and yet even though I don't think it

02:40

looks just like me it's still extremely

02:43

close to the reality for the amount of

02:45

information it was fed you can even try

02:48

it on yourself right now with the demo

02:49

no setup needed and it takes only a

02:52

minute or so click on the link in the

02:54

description run the program enter your

02:57

image which will not be saved by them

02:59

for privacy and here are the results in

03:02

this work they used a Celebi HQ data set

03:05

which is a large scale face attribute

03:07

data set with more than 200 thousand

03:09

celebrity images this could explain why

03:12

the faces look older and different than

03:15

non celebrity faces it will be

03:17

interesting to see a comparison with

03:19

another large data set using a different

03:22

segment of the population and this work

03:24

is of course generalizable and could be

03:26

of great use in fields like medicine

03:29

astronomy or satellite imagery where

03:32

sharp and high-resolution images are

03:34

difficult to have due to their cast or

03:36

hardware and memory limitations of

03:38

course this was just a simple overview

03:41

of the pulse photo op sampling algorithm

03:43

I strongly recommend to read the paper

03:46

and play with the code both are linked

03:48

in the description for more information

03:49

leave a like if you went this far in the

03:52

video and since there are over 90% of

03:55

you guys watching that are not

03:56

subscribed yet please consider

03:58

subscribing to the channel to not miss

03:59

any further news clearly explained