Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Deep Learning Image Manipulation
Search
Leszek Rybicki
May 18, 2017
Research
2
190
Deep Learning Image Manipulation
Illustrated guide to some image manipulation methods, with demonstration.
Leszek Rybicki
May 18, 2017
Tweet
Share
More Decks by Leszek Rybicki
See All by Leszek Rybicki
Let's talk about Fakes
lunardog
0
77
How to Patch Image Classifiers
lunardog
0
1.5k
Towards Realistic Predictors - EN
lunardog
0
1.3k
Towards Realistic Predictors
lunardog
1
1.7k
Deep Learning Hot Dog Detector
lunardog
0
200
Finding beans in burgers: paper reading notes
lunardog
0
1.1k
Kelner: Serve Your Models
lunardog
0
91
Image Analysis at Cookpad
lunardog
1
1.4k
Kelner: serve your models
lunardog
1
270
Other Decks in Research
See All in Research
音声処理ツールキットESPnetの現在と未来
kanbayashi1125
2
480
Generative Spoken Dialogue Language Modeling [対話論文読み会@電通大]
yuta0306
1
130
マルチモーダルLLMの応用動向の論文調査
masatoto
7
2.4k
Scaling Rectified Flow Transformers for High-Resolution Image Synthesis / Stable Diffusion 3
shunk031
0
150
プロシェアリング白書2024_PROSHARING_REPORT_2024
circulation
0
460
Accurate Method and Variable Tracking in Commit History
tsantalis
0
200
フルリモートワークでのスクラムのスケール
kmorita1111
2
960
Combating Misinformation in the age of LLMs
teacherpeterpan
0
110
Target trial emulationの概要
shuntaros
2
1.1k
「EBPMエコシステム」の可能性
daimoriwaki
0
190
Pathfinding for 10k agents
kei18
1
3k
[研究室用] 2038年問題研究の現状報告
ran350
0
250
Featured
See All Featured
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
272
12k
Web Components: a chance to create the future
zenorocha
304
41k
Done Done
chrislema
178
15k
Rails Girls Zürich Keynote
gr2m
91
13k
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
124
32k
Designing on Purpose - Digital PM Summit 2013
jponch
109
6.4k
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
113
18k
Let's Do A Bunch of Simple Stuff to Make Websites Faster
chriscoyier
501
140k
Designing Dashboards & Data Visualisations in Web Apps
destraynor
225
51k
Clear Off the Table
cherdarchuk
82
310k
Product Roadmaps are Hard
iamctodd
43
9.6k
Distributed Sagas: A Protocol for Coordinating Microservices
caitiem20
319
20k
Transcript
%FFQ-FBSOJOH *NBHF.BOJQVMBUJPO BOJMMVTUSBUFEHVJEF .-,JUDIFO
"CPVUNF w -FT[FL3ZCJDLJ w HJUIVC!MVOBSEPH w CPSOJO1PMBOE w .-3FTFBSDIFSBU$PPLQBE w
*MJLFOBUUP
DBSFFST!DPPLQBEDPN 8BOUUPXPSLXJUIVT
$POWPMVUJPOBM "SJUINFUJD OCIKE
*NBHFTUPGFBUVSFT
$POWPMVUJPO http://deeplearning.net/software/theano/tutorial/conv_arithmetic.html input output input output kernel
4USJEF http://deeplearning.net/software/theano/tutorial/conv_arithmetic.html 2px 2px 2px 2px
1BEEJOH http://deeplearning.net/software/theano/tutorial/conv_arithmetic.html 2px 2px
4USJEF QBEEJOH http://deeplearning.net/software/theano/tutorial/conv_arithmetic.html
5SBOTQPTFE http://deeplearning.net/software/theano/tutorial/conv_arithmetic.html simulated here with padding also called “deconvolution” “fractional
stride”
%PXOTBNQMJOH features or small resolution image convolutional layer or layers
RGB image input output
6QTBNQMJOH upsampling CNN layer or layers RGB image features or
small resolution image input output
&ODPEFS%FDPEFS D E image in Decoder Encoder image out feature
space
'VMMZ$POOFDUFE $MBTTJpFS approve loan reject class data or features also
called “Dense” layer
$//$MBTTJpFS food person plant other AlexNet, LeNet, VGG…
'PPE/FU ™ food not food
@teenybiscuit
None
@teenybiscuit
@teenybiscuit
@teenybiscuit
@teenybiscuit
@teenybiscuit
(FOFSBUJWF "EWFSTBSJBM /FUXPSLT
Generator Discriminator https://speakerdeck.com/lunardog/deep-convolutional-voight-kampf-test “Couple of bots studying for the Turing
Test”
Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks Alec
Radford, Luke Metz, Soumith Chintala (Submitted on 19 Nov 2015 (v1), last revised 7 Jan 2016 (this version, v2)) https://arxiv.org/abs/1511.06434
Generator Discriminator G MPPLTMFHJU UPUBMMZTIPQQFE D
G SFBM GBLF D D(G(noise)) ˠ real (FOFSBUPSUSBJOJOH Discriminator acts
as the teacher
G SFBM GBLF D SFBM GBLF D D(G(noise)) ˠ fake
D(photo) ˠ real %JTDSJNJOBUPSUSBJOJOH Generator provides negative examples
None
https://www.youtube.com/watch?v=rs3aI7bACGc ©Yota Ishida
Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks Alec
Radford, Luke Metz, Soumith Chintala (Submitted on 19 Nov 2015 (v1), last revised 7 Jan 2016 (this version, v2)) https://arxiv.org/abs/1511.06434
$POEJUJPOBM ("/T
G NBMF GFNBMF DIJME FMEFSMZ G(noise | conditions) $POEJUJPOBM(FOFSBUPS
SJHIU XSPOH NBMF GFNBMF DIJME FMEFSMZ D $POEJUJPOBM%JTDSJNJOBUPS
SJHIU XSPOH NBMF GFNBMF DIJME FMEFSMZ D SJHIU XSPOH NBMF
GFNBMF DIJME FMEFSMZ SJHIU XSPOH NBMF GFNBMF DIJME FMEFSMZ D D
SJHIU XSPOH D $POEJUJPOBM("/ https://arxiv.org/abs/1411.1784 Conditional Generative Adversarial Nets Mehdi
Mirza, Simon Osindero (Submitted on 6 Nov 2014) Generator Discriminator NBMF GFNBMF DIJME FMEFSMZ G NBMF GFNBMF DIJME FMEFSMZ same condition
G NBMF GFNBMF DIJME FMEFSMZ SJHIU XSPOH NBMF GFNBMF DIJME
FMEFSMZ D $POEJUJPOBM("/ Discriminator Generator
https://www.faceapp.com/ Disclaimer: FaceApp authors don’t disclose their method. This is
only my guess. It may have nothing to do with GANs. original
original https://www.faceapp.com/
https://www.faceapp.com/ original
"SUJTUJD4UZMF5SBOTGFS Improved!
https://prisma-ai.com/
https://prisma-ai.com/ https://prisma-ai.com/
https://prisma-ai.com/ https://prisma-ai.com/
https://prisma-ai.com/ https://prisma-ai.com/
https://arxiv.org/abs/1603.08155 transformation network loss network Gram matrices in feature space
pre-trained content image style image
“Gram matrices in feature space” https://en.wikipedia.org/wiki/Gramian_matrix
https://www.youtube.com/watch?v=xVJwwWQlQ1o
$ZDMF("/
https://github.com/junyanz/CycleGAN
https://github.com/junyanz/CycleGAN
https://github.com/junyanz/CycleGAN
(FOFSBUPS transformation network https://arxiv.org/abs/1603.08155
GBLF IPSTF GBLF IPSTF … %JTDSJNJOBUPS fully convolutional judges patches
of the input image https://arxiv.org/abs/1603.08155
"EWFSTBSJBM-PTT X F G Y GBLF [FCSB GBLF [FCSB …
GBLF IPSTF GBLF IPSTF … X(F(horse)) ˠ classify as zebra Y(F(zebra)) ˠ classify as horse
$ZDMF-PTT G F G(F(image))ˠ the same image F G F(G(image))ˠ
the same image
https://www.youtube.com/watch?v=9reHvktowLY
5IF&OE