SlideShare a Scribd company logo
1 of 46
Download to read offline
Image Style Transfer Using
Convolutional Neural Networks (2016)
(a.k.a. Neural Style Transfer)
Oct, 2018
Sooyoung Moon
2
sabjil.jpg julgyu.jpg
Content Image Style Image
Example
3
julgyuesabjil.jpg
New Image
Example
4
Abstract
์ด์ „์—๋Š”..
Rendering the semantic content of an image in different styles is a difficult image processing task.
์—ฌ๋Ÿฌ ๋‹ค๋ฅธ style image์— content๋ฅผ ๋ Œ๋”๋งํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ค์šด ์ผ์ด๋‹ค. ์šฐ๋ฆฌ๋Š” ์–ด๋ ค์šด ์ผ์„
ํ•ด๋ƒˆ๋‹ค!!
Arguably, a major limiting factor for previous approaches has been the lack of image
representations that explicitly represent semantic information and, thus, allow to separate image
content from style.
์ด์ „์—๋Š” ์˜๋ฏธ ์žˆ๋Š” ์ •๋ณด(contents)๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„ํ•˜๋Š” ๊ฒƒ์ด ๋ถ€์กฑํ–ˆ๋‹ค. ๊ณ ๋กœ,
์Šคํƒ€์ผ๋กœ๋ถ€ํ„ฐ ์ด๋ฏธ์ง€ ์ปจํ…์ธ ๋ฅผ ๋ถ„๋ฆฌํ•˜๊ฒŒ ๊ฐ€๋Šฅํ•˜๋Š” ๊ฒƒ์— ํ•œ๊ณ„๊ฐ€ ์žˆ์—ˆ๋‹ค. ์–ด๋–ค ๋ฐฉ๋ฒ•์„ ์ผ๊ธธ๋ž˜?
5
The algorithm allows us to produce new images of high perceptual quality that combine the
content of an arbitrary photograph with the appearance of numerous well-known artworks.
์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์ž˜ ์•Œ๋ ค์ง„ ์˜ˆ์ˆ ์ž‘ํ’ˆ์˜ ์Šคํƒ€์ผ๊ณผ ์ž„์˜์˜ ์‚ฌ์ง„์˜ ์ปจํ…์ธ ๋ฅผ ์กฐํ•ฉํ•˜์—ฌ ์ƒˆ๋กœ์šด
๊ณ ํ€„๋ฆฌํ‹ฐ์˜ ์ด๋ฏธ์ง€๋ฅผ ๋งŒ๋“ค์–ด ๋‚ธ๋‹ค.
Our results provide new insights into the deep image representations learned by Convolutional
Neural Networks and demonstrate their potential for high level image synthesis and
manipulation.
์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋ฌผ์€ CNN์— ์˜ํ•œ deep image representations์— ๋Œ€ํ•ด ์ƒˆ๋กœ์šด ์ธ์‹ธ์ดํŠธ๋ฅผ ์ œ๊ณต
ํ•œ๋‹ค. ๊ทธ๋ฆฌ๊ณ  high level image synthesis & manipulation ์— ๋Œ€ํ•œ ํฌํ…์…œ์„ ์ฆ๋ช…ํ•œ๋‹ค.
Abstract
์ด์ œ๋Š”..
6
Introduction
์ด์ „์—๋Š”..
Texture transfering์€ ์˜ˆ์ „์—๋„ ์—ฌ๋Ÿฌ ๋ฐฉ๋ฒ•๋“ค์ด ์žˆ์—ˆ๋‹ค.
์ฃผ๋กœ texture transfer algorithm์˜ ๋ฒ ์ด์ง์€ non-parametric methods์˜€๊ณ , ํƒ€๊ฒŸ ์ด๋ฏธ์ง€์˜
structure์„ ๋ณด์กดํ•˜๋Š” ๋ฐฉ๋ฒ•๋“ค์ด ๋‹ฌ๋ž๋‹ค.
์˜ˆ๋ฅผ ๋“ค๋ฉด,
- Efros&Freeman: correspondence map [link]
- Image Analogies (2001, Hertzman et al) [link]
- Fast Texture Transfer (2001, Ashikhmin) [link]
- Directional Texture Transfer (2010, Lee et al.) [link]
7
Image Quilting for Texture Synthesis and Transfer (2001, Efros & Freeman) [link]
์–ผ๊ตด์— ๋ฐฅํ’€ ๋ถ™์ด๊ธฐ ๋ผ๋˜๊ฐ€..
์•„๋ จ โ€ฆ
8
Image Quilting for Texture Synthesis and Transfer (2001, Efros & Freeman) [link]
์–ผ๊ตด์— ๋ฐฅํ’€ ๋ถ™์ด๊ธฐ ๋ผ๋˜๊ฐ€..
9
Image Analogies (2001, Hertzman et al) [link]
์—ฌ๊ธฐ๋Š” ๊ฐ•๋ฌผ ์žฌ์ฐฝ์กฐ..
10
Fast Texture Transfer (2001, Ashikhmin) [link]
์˜ˆ์œ ๊ฝƒ ์‚ฌ์ง„์— ํ•„ํ„ฐ ๋ฟŒ๋ฆฌ๊ธฐ..
๊ธ‰๋…ธํ™”๊ฐ€ ์˜จ ๊ฒƒ ๊ฐ™์€ ๊ฒƒ์€ ๊ธฐ๋ถ„ ํƒ“์ด์•ผ....
11
Directional Texture Transfer (2010, Lee et al.) [link]
๋ˆ์ด ์—†์–ด์„œ ์—ด์–ด๋ณด์ง€ ๋ชปํ–ˆ์Œ..
12
Directional Texture Transfer (2010, Lee et al.) [link]
๋ˆ์ด ์—†์–ด์„œ ์—ด์–ด๋ณด์ง€ ๋ชปํ–ˆ์Œ..
๋‚œโ€ฆใ„ฑ ใ…๋”โ€ฆ
๋ˆˆ๋ฌผ์„ ํ˜๋ฆฐ ใ„ท ใ…โ€ฆ.
13
๊ทธ๋ž˜์„œ ๊ตฌ๊ธ€๋ง..
์ž๋ž‘์Šค๋Ÿฌ์šด ํ•œ๊ตญ์ธโ€ฆ!
Directional Texture Transfer (2010, Lee et al.) [link]
14
Introduction
์ด์ „์—๋Š”..
Texture transfering์€ ์˜ˆ์ „์—๋„ ์—ฌ๋Ÿฌ ๋ฐฉ๋ฒ•๋“ค์ด ์žˆ์—ˆ๋‹ค.
๊ทธ๋Ÿฌ๋‚˜ ํ•œ๊ณ„๊ฐ€ ์žˆ์—ˆ๋‹ค: Texture transfer๋ฅผ ํ•˜๊ธฐ ์œ„ํ•ด ์˜ค๋กœ์ง€ ํƒ€๊ฒŸ ์ด๋ฏธ์ง€์˜ Low-level feature๋“ค๋งŒ
์‚ฌ์šฉํ–ˆ๋‹ค.
์ด์ƒ์ ์œผ๋กœ ๋ณด๋ฉด, Content image๋กœ๋ถ€ํ„ฐ ์ปจํ…์ธ ๋ฅผ ์ถ”์ถœํ•œ ๋‹ค์Œ Style image์˜ style์— ๋ Œ๋”๋งํ•˜๋Š”
Texture transfer ์ ˆ์ฐจ๊ฐ€ ์ข‹์•„๋ณด์ธ๋‹ค.
15
Introduction
๊ทธ๋Ÿฌ๋ฏ€๋กœ ์ด๋ฏธ์ง€์˜ representations ์ฐพ๋Š” ๊ฒƒ์ด ์šฐ์„ ์ด ๋˜์–ด์•ผํ•œ๋‹ค. semantic image content์˜
variation์ด๋ž‘ ๊ทธ๋ฆผ์— ๋‚˜ํƒ€๋‚ผ ์Šคํƒ€์ผ์— ๋Œ€ํ•ด์„œ ๊ฐ๊ฐ ๋…๋ฆฝ์ ์œผ๋กœ ๋ชจ๋ธ๋งํ•˜๋Š” ๊ฒƒ์ด ์ด๋ฏธ์ง€์˜
representation์ด๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค.
์ด๋Ÿฌํ•œ factorized representations๋Š” ์ด๋ฏธ ์ด์ „์— ์ด๋ฃจ์–ด์ง„ ์  ์žˆ๋‹ค. ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ œํ•œ๋œ
์ด๋ฏธ์ง€์—์„œ ์ˆ˜ํ–‰๋˜์—ˆ๋‹ค. ๋‹ค๋ฅธ ์กฐ๋ช… ์•„๋ž˜์˜ ์–ผ๊ตด๋“ค, ๋‹ค๋ฅธ ํฐํŠธ ์Šคํƒ€์ผ์˜ ์ฒ ์ž๋“ค, ์†์œผ๋กœ
์“ฐ์—ฌ์ง„ ์ˆซ์ž๋“ค, ๋ฒˆ์ง€์ˆ˜ ๋“ฑ์ด ๊ทธ ์˜ˆ์‹œ๋‹ค.
16
Introduction
์Šคํƒ€์ผ๋กœ๋ถ€ํ„ฐ ์ปจํ…์ธ  ๋ถ„๋ฆฌ๋Š” ์—ฌ์ „ํžˆ ์–ด๋ ค์šด ๋ฌธ์ œ๋‹ค. ํ•˜์ง€๋งŒ Deep CNN์˜ ๋ฐœ์ „์€ ๊ฐ•๋ ฅํ•œ
์ปดํ“จํ„ฐ ๋น„์ „ ์‹œ์Šคํ…œ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ด์คฌ๋‹ค. ๋” ๊ตฌ์ฒด์ ์œผ๋กœ ๋งํ•˜์ž๋ฉด, Natural images์˜ high-
level ์˜๋ฏธ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ฒƒ์„ ํ•™์Šต์‹œํ‚จ ์ปดํ“จํ„ฐ ๋น„์ „ ์‹œ์Šคํ…œ์ด ๊ฐ€๋Šฅํ•ด ์กŒ๋‹ค.
์ถฉ๋ถ„ํ•œ ์–‘์˜ ๋ ˆ์ด๋ธ”๋œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต์‹œํ‚จ CNN์€ ๋ฐ์ดํ„ฐ์…‹๋“ค ์ „๋ฐ˜์ ์œผ๋กœ ์ผ๋ฐ˜ํ™”๋˜์–ด ์žˆ๋Š”
feature representations๋“ค์—์„œ high-level ์ด๋ฏธ์ง€ ์ปจํ…์ธ ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ฒƒ์„ ํ•™์Šตํ•œ๋‹ค.
17
Introduction
์ด ๋…ผ๋ฌธ์—์„œ๋Š” deep CNN์˜๋กœ ํ•™์Šต๋œ generic feature representations๊ฐ€ natural images๋“ค์˜
์ปจํ…์ธ ์™€ ์Šคํƒ€์ผ์„ ๋…๋ฆฝ์ ์œผ๋กœ ์กฐ์ž‘ํ•˜๊ณ  ๋…๋ฆฝ์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์— ์‚ฌ์šฉ๋˜๋Š”์ง€ ๋ณด์—ฌ์ค€๋‹ค.
๋ชจ๋ธ์ด ๋”ฅํ•˜๋‹ค๋ณด๋‹ˆ ์‹ฑ๊ธ€ ๋‰ด๋Ÿด๋„ท์—์„œ ์žˆ์—ˆ๋˜ ์˜ตํ‹ฐ๋งˆ์ด์ œ์ด์…˜ ๋ฌธ์ œ๋ฅผ ์ค„์ผ ์ˆ˜ ์žˆ๋‹ค. ์ƒˆ๋กœ์šด
์ด๋ฏธ์ง€๋Š” ์–ด๋–ป๊ฒŒ ๋งŒ๋“ค์–ด์ง€๋ƒ ํ•˜๋ฉด ์ƒ˜ํ”Œ์ด๋ฏธ์ง€์˜ feature representations์™€ ์–ด๋– ํ•œ pre-
image๋ฅผ ๋งค์นญ ์‹œํ‚ค๋ฉด์„œ pre-image์˜ ๋ชจ์Šต์„ ์ฐพ์•„๋‚˜๊ฐ€๋Š” ๋ฐฉ๋ฒ•์ด๋‹ค.
์‹ค์ œ๋กœ, ์šฐ๋ฆฌ์˜ ์Šคํƒ€์ผ ํŠธ๋žœ์Šคํผ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ โ€˜image representationsโ€™์˜ ๋ฐ˜๋Œ€ ๋ฐฉ๋ฒ•์„ ์“ฐ๋Š”
CNN์ด ๋ฒ ์ด์Šค์ธ model์„ ์“ด๋‹ค.
18
Deep image representations
Pre-trained 19 layer VGG network ์‚ฌ์šฉ
์ถœ์ฒ˜: https://www.youtube.com/watch?v=fIW8fI2Xb_k
19
Deep image representations
์ถœ์ฒ˜: https://www.youtube.com/watch?v=fIW8fI2Xb_k
20
๋’ค๋กœ ๊ฐˆ์ˆ˜๋ก ๋””ํ…Œ์ผ ํ”ฝ์…€ ์ •๋ณด๊ฐ€
๋กœ์Šค๋œ๋‹ค.
์„œ๋กœ ๋‹ค๋ฅธ ๋ ˆ์ด์–ด๋“ค๊ฐ„์˜
์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ถ„์„ํ•œ๋‹ค. conv1 /
conv1,conv2 / conv1, conv2, conv3
/ conv1,conv2,conv3,con4 /
conv1,conv2,conv3, conv4, conv5
e: ๋งˆ์นจ๋‚ด global arrangement ๋Š”
์‚ฌ๋ผ์ ธ๋„ ์ฃผ์–ด์ง„ ์ด๋ฏธ์ง€์˜
์Šคํƒ€์ผ์ด ์ž˜ ์ƒ์„ฑ๋œ๋‹ค.
21
2.1 Style representation
22์ถœ์ฒ˜: https://www.youtube.com/watch?v=fIW8fI2Xb_k
23
Content representation
โƒ—" : generated image
โƒ—# : original image
$% : original image์˜ feature representation in layer &
'%
: generated image์˜ feature representation in layer &
์œ„์˜ ๋กœ์Šค ํ•จ์ˆ˜๋ฅผ ํ†ตํ•ด gradient descen๋ฅผ ํ†ตํ•ด โƒ—"๊ฐ€ โƒ—#๊ฐ€ ๋˜๊ฒŒ๋” ํ•œ๋‹ค.
24
Style representation
Layer ๋“ค ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ถ„์„ํ•˜๋Š” ํ‹€์ธ ๊ทธ๋žจ ๋งคํŠธ๋ฆญ์Šค๋ฅผ ์‚ฌ์šฉ
์ถœ์ฒ˜: ๋ณด์ฐฌ๋‹˜์˜ ์Šฌ๋ผ์ด๋“œ(CycleGAN) Gram Matrix ํ™์ •๋ชจ ๊ต์ˆ˜๋‹˜ ์œ ํˆฌ๋ธŒ ๋ฐ”๋กœ๊ฐ€๊ธฐ (ํด๋ฆญ!)
25
Results
26
Results
27
Results
28
Trade-off between content and style matching
29
Effect of different layers of the Convolutional Neural Network
30
Initialisation of gradient descent
31
Photorealistic style transfer
32
Discussion
33
๊ฝค ์ž˜ํ•œ ๊ฒƒ ๊ฐ™์ง€๋งŒ ๊ทธ๋ž˜๋„ ํ•œ๊ณ„์ ์ด ์žˆ๋‹ค.
Discussion
34
LIMITATION 1
Resolution of the synthesized images.
The speed of the synthesis procedure depends heavily on image resolution.
Discussion
35
LIMITATION 2
Noise.
While this is less of an is- sue in the artistic style transfer, the problem becomes more apparent
when both, content and style images, are photographs.
Discussion
36
LIMITATION 2Discussion
์ œ๊ฐ€ ์ง์ ‘ ํ•œ๋ฒˆ ํ…Œ์ŠคํŠธ ํ•ด๋ณด๊ฒ ์๋‹ˆ๋‹ค..
์‚ฝ์งˆ ํ•  ์ค€๋น„๊ฐ€ ๋˜์–ด์žˆ๋Š” ๋ฌธ์ˆ˜์˜์”จ ์ด๋ฏธ ์‚ฝ์งˆ ์ค‘์ธ ์•„์ €์”จ
37
LIMITATION 2Discussion
์‚ฝ์งˆ ํ•  ์ค€๋น„๊ฐ€ ๋˜์–ด์žˆ๋Š” ๋ฌธ์ˆ˜์˜์”จ ์ด๋ฏธ ์‚ฝ์งˆ ์ค‘์ธ ์•„์ €์”จ
LIMITATION 1์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ๋น ๋ฅธ ํ•™์Šต์„ ์œ„ํ•ด ๋””๋ฉ˜์…˜์€ 128*128๋กœ ๋ฐ”๊ฟจ์Šต๋‹ˆ๋‹ค...
38
LIMITATION 2Discussion
๋ฉ”๋งˆ๋ฅธ ๋•…์— ๊ฐ•๋ฌผ์ด ํ๋ฅด๊ฒŒ ๋˜๊ณ  ํ•˜๋Š˜์—๋Š” ํ’€์ด ์ž๋ผ๋‚˜๋‹ˆ
39
LIMITATION 2Discussion
์›๋ž˜๋ถ€ํ„ฐ ํ™”์งˆ์ด ์•ˆ์ข‹์•„์„œ ๋…ธ์ด์ฆˆ๊ฐ€ ์ƒ๊ธด์ง€ ์ž˜ ๋ชจ๋ฅด๊ฒ ์Šตโ€ฆ
40
LIMITATION 2Discussion
์˜ค๋ฆฌ์ง€๋‚  ์ด๋ฏธ์ง€๋ž‘ ๋น„๊ตํ•ด๋ณด๋ฉด ๋…ธ์ด์ฆˆ๊ฐ€ ์ข€ ์ƒ๊ธด ๊ฒƒ ๊ฐ™๊ธฐ๋„โ€ฆ
41
LIMITATION 2
However, the noise is very characteristic and appears to resemble the filters of units in the
network.
Discussion
42
LIMITATION 3
The separation of image content from style is not necessarily a well defined problem.
This is mostly because it is not clear what exactly defines the style of an image.
It might be the brush strokes in a painting, the colour map, certain dominant forms and shapes,
but also the composition of a scene and the choice of the subject of the image and probably it is
a mixture of all of them and many more.
Discussion
43
LIMITATION 3Discussion
์˜ˆ๋ฅผ ๋“ค๋ฉด,
๋‚ด๊ฐ€ ๋ˆˆ์น ์™์‹ฑ์„ ํ•˜๊ณ  ์‹ถ์–ด์„œ
๋ชจ๋‚˜๋ฆฌ์ž ์Šคํƒ€์ผ์„ ๊ฐ€์ ธ๋‹ค๊ฐ€
์“ฐ๊ฒ ๋‹ค๊ณ  ํ•ด์„œ ๋‚ด ๋ˆˆ์น์ด ์‚ฌ๋ผ์ง€๋Š”
๊ฒƒ์ด ์•„๋‹ˆ๋‹ค..
๋‚ด ์ •๋ฉด์˜ ๋ชจ์Šต์ด ๋งˆ์Œ์— ์•ˆ ๋“ค์–ด์„œ
์ธก๋ฉด์„ ๊ธฐ๋Œ€ํ•˜๊ณ  ์ €๋Ÿฐ ๋‚จ์ž ์‚ฌ์ง„
์Šคํƒ€์ผ์„ ๊ฐ€์ ธ๋‹ค ์“ด๋‹ค๊ณ  ํ•ด์„œ ๋‚ด๊ฐ€
์–ผ๊ตด์„ ๋Œ๋ฆฌ๊ณ  ์ฐ์€ ์‚ฌ์ง„์ด ํ•ฉ์„ฑ๋˜๋Š”
๊ฒƒ์ด ์•„๋‹ˆ๋‹ค..
์‚ฌ๊ธฐ ์‚ฌ์ง„ ๋“ฑ์žฅ ใ…Ž
44
LIMITATION 3
In our work we consider style transfer to be successful if the generated image โ€˜looks likeโ€™ the style
image but shows the objects and scenery of the content image.
ํ˜„์‹ค๊ณผ ํƒ€ํ˜‘ํ•  ์‹œ๊ฐ„โ€ฆ
Discussion
45
Nevertheless,
์ƒ๋ฌผํ•™์  ์‹œ๊ฐ ์ฒด๊ณ„๊ฐ€ ์ˆ˜ํ–‰ํ•˜๋Š” ์ผ๋“ค ์ค‘ ํ•˜๋‚˜๋ฅผ ํ•ด๋‚ผ ์ˆ˜ ์žˆ๋„๋ก ํ›ˆ๋ จ๋œ Neural Network๊ฐ€
์–ด๋Š ์ •๋„๋Š” style๋กœ๋ถ€ํ„ฐ contents๋ฅผ ์ž๋™์œผ๋กœ ๋ถ„๋ฆฌํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ด ๊ต‰์žฅํžˆ fascinatingํ•˜๋‹ค.
Discussion
46
๋งˆ๋ฌด๋ฆฌ
์ตœ์ ํ™”๋œ ANN์˜ ์„ฑ๋Šฅ๊ณผ ์‚ฌ๋žŒ์˜ ๋น„์ „์ด ์„œ๋กœ ์ƒ๋‹นํžˆ ๋‹ฎ์•˜๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•ด๋ณด๋ฉด
์‚ฌ๋žŒ๋„ ์—ญ์‹œ contents ๋ฅผ style๋กœ๋ถ€ํ„ฐ ์ถ”๋ก ํ•ด๋‚ด๋Š” ๋Šฅ๋ ฅ์ด ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋ ‡๊ธฐ
๋•Œ๋ฌธ์— ์˜ˆ์ˆ ์„ ์ฐฝ์กฐํ•˜๊ณ  ์ฆ๊ธธ ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์ด๋‹ค. ์ฆ‰, ์šฐ๋ฆฌ์˜ ์‹œ๊ฐ ์ฒด๊ณ„์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์— ๋Œ€ํ•œ
ํ™•์‹คํ•œ ํŠน์„ฑ์„ ์œ ์ถ”ํ•  ์ˆ˜ ์žˆ๋‹ค. ์•„๋ฆ„๋‹ค์šด ๋งˆ๋ฌด๋ฆฌ
Discussion

More Related Content

Featured

2024 State of Marketing Report โ€“ by Hubspot
2024 State of Marketing Report โ€“ by Hubspot2024 State of Marketing Report โ€“ by Hubspot
2024 State of Marketing Report โ€“ by HubspotMarius Sescu
ย 
Everything You Need To Know About ChatGPT
Everything You Need To Know About ChatGPTEverything You Need To Know About ChatGPT
Everything You Need To Know About ChatGPTExpeed Software
ย 
Product Design Trends in 2024 | Teenage Engineerings
Product Design Trends in 2024 | Teenage EngineeringsProduct Design Trends in 2024 | Teenage Engineerings
Product Design Trends in 2024 | Teenage EngineeringsPixeldarts
ย 
How Race, Age and Gender Shape Attitudes Towards Mental Health
How Race, Age and Gender Shape Attitudes Towards Mental HealthHow Race, Age and Gender Shape Attitudes Towards Mental Health
How Race, Age and Gender Shape Attitudes Towards Mental HealthThinkNow
ย 
AI Trends in Creative Operations 2024 by Artwork Flow.pdf
AI Trends in Creative Operations 2024 by Artwork Flow.pdfAI Trends in Creative Operations 2024 by Artwork Flow.pdf
AI Trends in Creative Operations 2024 by Artwork Flow.pdfmarketingartwork
ย 
PEPSICO Presentation to CAGNY Conference Feb 2024
PEPSICO Presentation to CAGNY Conference Feb 2024PEPSICO Presentation to CAGNY Conference Feb 2024
PEPSICO Presentation to CAGNY Conference Feb 2024Neil Kimberley
ย 
Content Methodology: A Best Practices Report (Webinar)
Content Methodology: A Best Practices Report (Webinar)Content Methodology: A Best Practices Report (Webinar)
Content Methodology: A Best Practices Report (Webinar)contently
ย 
How to Prepare For a Successful Job Search for 2024
How to Prepare For a Successful Job Search for 2024How to Prepare For a Successful Job Search for 2024
How to Prepare For a Successful Job Search for 2024Albert Qian
ย 
Social Media Marketing Trends 2024 // The Global Indie Insights
Social Media Marketing Trends 2024 // The Global Indie InsightsSocial Media Marketing Trends 2024 // The Global Indie Insights
Social Media Marketing Trends 2024 // The Global Indie InsightsKurio // The Social Media Age(ncy)
ย 
Trends In Paid Search: Navigating The Digital Landscape In 2024
Trends In Paid Search: Navigating The Digital Landscape In 2024Trends In Paid Search: Navigating The Digital Landscape In 2024
Trends In Paid Search: Navigating The Digital Landscape In 2024Search Engine Journal
ย 
5 Public speaking tips from TED - Visualized summary
5 Public speaking tips from TED - Visualized summary5 Public speaking tips from TED - Visualized summary
5 Public speaking tips from TED - Visualized summarySpeakerHub
ย 
ChatGPT and the Future of Work - Clark Boyd
ChatGPT and the Future of Work - Clark Boyd ChatGPT and the Future of Work - Clark Boyd
ChatGPT and the Future of Work - Clark Boyd Clark Boyd
ย 
Getting into the tech field. what next
Getting into the tech field. what next Getting into the tech field. what next
Getting into the tech field. what next Tessa Mero
ย 
Google's Just Not That Into You: Understanding Core Updates & Search Intent
Google's Just Not That Into You: Understanding Core Updates & Search IntentGoogle's Just Not That Into You: Understanding Core Updates & Search Intent
Google's Just Not That Into You: Understanding Core Updates & Search IntentLily Ray
ย 
Introduction to Data Science
Introduction to Data ScienceIntroduction to Data Science
Introduction to Data ScienceChristy Abraham Joy
ย 
Time Management & Productivity - Best Practices
Time Management & Productivity -  Best PracticesTime Management & Productivity -  Best Practices
Time Management & Productivity - Best PracticesVit Horky
ย 
The six step guide to practical project management
The six step guide to practical project managementThe six step guide to practical project management
The six step guide to practical project managementMindGenius
ย 
Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...
Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...
Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...RachelPearson36
ย 

Featured (20)

2024 State of Marketing Report โ€“ by Hubspot
2024 State of Marketing Report โ€“ by Hubspot2024 State of Marketing Report โ€“ by Hubspot
2024 State of Marketing Report โ€“ by Hubspot
ย 
Everything You Need To Know About ChatGPT
Everything You Need To Know About ChatGPTEverything You Need To Know About ChatGPT
Everything You Need To Know About ChatGPT
ย 
Product Design Trends in 2024 | Teenage Engineerings
Product Design Trends in 2024 | Teenage EngineeringsProduct Design Trends in 2024 | Teenage Engineerings
Product Design Trends in 2024 | Teenage Engineerings
ย 
How Race, Age and Gender Shape Attitudes Towards Mental Health
How Race, Age and Gender Shape Attitudes Towards Mental HealthHow Race, Age and Gender Shape Attitudes Towards Mental Health
How Race, Age and Gender Shape Attitudes Towards Mental Health
ย 
AI Trends in Creative Operations 2024 by Artwork Flow.pdf
AI Trends in Creative Operations 2024 by Artwork Flow.pdfAI Trends in Creative Operations 2024 by Artwork Flow.pdf
AI Trends in Creative Operations 2024 by Artwork Flow.pdf
ย 
Skeleton Culture Code
Skeleton Culture CodeSkeleton Culture Code
Skeleton Culture Code
ย 
PEPSICO Presentation to CAGNY Conference Feb 2024
PEPSICO Presentation to CAGNY Conference Feb 2024PEPSICO Presentation to CAGNY Conference Feb 2024
PEPSICO Presentation to CAGNY Conference Feb 2024
ย 
Content Methodology: A Best Practices Report (Webinar)
Content Methodology: A Best Practices Report (Webinar)Content Methodology: A Best Practices Report (Webinar)
Content Methodology: A Best Practices Report (Webinar)
ย 
How to Prepare For a Successful Job Search for 2024
How to Prepare For a Successful Job Search for 2024How to Prepare For a Successful Job Search for 2024
How to Prepare For a Successful Job Search for 2024
ย 
Social Media Marketing Trends 2024 // The Global Indie Insights
Social Media Marketing Trends 2024 // The Global Indie InsightsSocial Media Marketing Trends 2024 // The Global Indie Insights
Social Media Marketing Trends 2024 // The Global Indie Insights
ย 
Trends In Paid Search: Navigating The Digital Landscape In 2024
Trends In Paid Search: Navigating The Digital Landscape In 2024Trends In Paid Search: Navigating The Digital Landscape In 2024
Trends In Paid Search: Navigating The Digital Landscape In 2024
ย 
5 Public speaking tips from TED - Visualized summary
5 Public speaking tips from TED - Visualized summary5 Public speaking tips from TED - Visualized summary
5 Public speaking tips from TED - Visualized summary
ย 
ChatGPT and the Future of Work - Clark Boyd
ChatGPT and the Future of Work - Clark Boyd ChatGPT and the Future of Work - Clark Boyd
ChatGPT and the Future of Work - Clark Boyd
ย 
Getting into the tech field. what next
Getting into the tech field. what next Getting into the tech field. what next
Getting into the tech field. what next
ย 
Google's Just Not That Into You: Understanding Core Updates & Search Intent
Google's Just Not That Into You: Understanding Core Updates & Search IntentGoogle's Just Not That Into You: Understanding Core Updates & Search Intent
Google's Just Not That Into You: Understanding Core Updates & Search Intent
ย 
How to have difficult conversations
How to have difficult conversations How to have difficult conversations
How to have difficult conversations
ย 
Introduction to Data Science
Introduction to Data ScienceIntroduction to Data Science
Introduction to Data Science
ย 
Time Management & Productivity - Best Practices
Time Management & Productivity -  Best PracticesTime Management & Productivity -  Best Practices
Time Management & Productivity - Best Practices
ย 
The six step guide to practical project management
The six step guide to practical project managementThe six step guide to practical project management
The six step guide to practical project management
ย 
Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...
Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...
Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...
ย 

Neural-style-transfer

  • 1. Image Style Transfer Using Convolutional Neural Networks (2016) (a.k.a. Neural Style Transfer) Oct, 2018 Sooyoung Moon
  • 4. 4 Abstract ์ด์ „์—๋Š”.. Rendering the semantic content of an image in different styles is a difficult image processing task. ์—ฌ๋Ÿฌ ๋‹ค๋ฅธ style image์— content๋ฅผ ๋ Œ๋”๋งํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ค์šด ์ผ์ด๋‹ค. ์šฐ๋ฆฌ๋Š” ์–ด๋ ค์šด ์ผ์„ ํ•ด๋ƒˆ๋‹ค!! Arguably, a major limiting factor for previous approaches has been the lack of image representations that explicitly represent semantic information and, thus, allow to separate image content from style. ์ด์ „์—๋Š” ์˜๋ฏธ ์žˆ๋Š” ์ •๋ณด(contents)๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„ํ•˜๋Š” ๊ฒƒ์ด ๋ถ€์กฑํ–ˆ๋‹ค. ๊ณ ๋กœ, ์Šคํƒ€์ผ๋กœ๋ถ€ํ„ฐ ์ด๋ฏธ์ง€ ์ปจํ…์ธ ๋ฅผ ๋ถ„๋ฆฌํ•˜๊ฒŒ ๊ฐ€๋Šฅํ•˜๋Š” ๊ฒƒ์— ํ•œ๊ณ„๊ฐ€ ์žˆ์—ˆ๋‹ค. ์–ด๋–ค ๋ฐฉ๋ฒ•์„ ์ผ๊ธธ๋ž˜?
  • 5. 5 The algorithm allows us to produce new images of high perceptual quality that combine the content of an arbitrary photograph with the appearance of numerous well-known artworks. ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์ž˜ ์•Œ๋ ค์ง„ ์˜ˆ์ˆ ์ž‘ํ’ˆ์˜ ์Šคํƒ€์ผ๊ณผ ์ž„์˜์˜ ์‚ฌ์ง„์˜ ์ปจํ…์ธ ๋ฅผ ์กฐํ•ฉํ•˜์—ฌ ์ƒˆ๋กœ์šด ๊ณ ํ€„๋ฆฌํ‹ฐ์˜ ์ด๋ฏธ์ง€๋ฅผ ๋งŒ๋“ค์–ด ๋‚ธ๋‹ค. Our results provide new insights into the deep image representations learned by Convolutional Neural Networks and demonstrate their potential for high level image synthesis and manipulation. ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋ฌผ์€ CNN์— ์˜ํ•œ deep image representations์— ๋Œ€ํ•ด ์ƒˆ๋กœ์šด ์ธ์‹ธ์ดํŠธ๋ฅผ ์ œ๊ณต ํ•œ๋‹ค. ๊ทธ๋ฆฌ๊ณ  high level image synthesis & manipulation ์— ๋Œ€ํ•œ ํฌํ…์…œ์„ ์ฆ๋ช…ํ•œ๋‹ค. Abstract ์ด์ œ๋Š”..
  • 6. 6 Introduction ์ด์ „์—๋Š”.. Texture transfering์€ ์˜ˆ์ „์—๋„ ์—ฌ๋Ÿฌ ๋ฐฉ๋ฒ•๋“ค์ด ์žˆ์—ˆ๋‹ค. ์ฃผ๋กœ texture transfer algorithm์˜ ๋ฒ ์ด์ง์€ non-parametric methods์˜€๊ณ , ํƒ€๊ฒŸ ์ด๋ฏธ์ง€์˜ structure์„ ๋ณด์กดํ•˜๋Š” ๋ฐฉ๋ฒ•๋“ค์ด ๋‹ฌ๋ž๋‹ค. ์˜ˆ๋ฅผ ๋“ค๋ฉด, - Efros&Freeman: correspondence map [link] - Image Analogies (2001, Hertzman et al) [link] - Fast Texture Transfer (2001, Ashikhmin) [link] - Directional Texture Transfer (2010, Lee et al.) [link]
  • 7. 7 Image Quilting for Texture Synthesis and Transfer (2001, Efros & Freeman) [link] ์–ผ๊ตด์— ๋ฐฅํ’€ ๋ถ™์ด๊ธฐ ๋ผ๋˜๊ฐ€.. ์•„๋ จ โ€ฆ
  • 8. 8 Image Quilting for Texture Synthesis and Transfer (2001, Efros & Freeman) [link] ์–ผ๊ตด์— ๋ฐฅํ’€ ๋ถ™์ด๊ธฐ ๋ผ๋˜๊ฐ€..
  • 9. 9 Image Analogies (2001, Hertzman et al) [link] ์—ฌ๊ธฐ๋Š” ๊ฐ•๋ฌผ ์žฌ์ฐฝ์กฐ..
  • 10. 10 Fast Texture Transfer (2001, Ashikhmin) [link] ์˜ˆ์œ ๊ฝƒ ์‚ฌ์ง„์— ํ•„ํ„ฐ ๋ฟŒ๋ฆฌ๊ธฐ.. ๊ธ‰๋…ธํ™”๊ฐ€ ์˜จ ๊ฒƒ ๊ฐ™์€ ๊ฒƒ์€ ๊ธฐ๋ถ„ ํƒ“์ด์•ผ....
  • 11. 11 Directional Texture Transfer (2010, Lee et al.) [link] ๋ˆ์ด ์—†์–ด์„œ ์—ด์–ด๋ณด์ง€ ๋ชปํ–ˆ์Œ..
  • 12. 12 Directional Texture Transfer (2010, Lee et al.) [link] ๋ˆ์ด ์—†์–ด์„œ ์—ด์–ด๋ณด์ง€ ๋ชปํ–ˆ์Œ.. ๋‚œโ€ฆใ„ฑ ใ…๋”โ€ฆ ๋ˆˆ๋ฌผ์„ ํ˜๋ฆฐ ใ„ท ใ…โ€ฆ.
  • 14. 14 Introduction ์ด์ „์—๋Š”.. Texture transfering์€ ์˜ˆ์ „์—๋„ ์—ฌ๋Ÿฌ ๋ฐฉ๋ฒ•๋“ค์ด ์žˆ์—ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ•œ๊ณ„๊ฐ€ ์žˆ์—ˆ๋‹ค: Texture transfer๋ฅผ ํ•˜๊ธฐ ์œ„ํ•ด ์˜ค๋กœ์ง€ ํƒ€๊ฒŸ ์ด๋ฏธ์ง€์˜ Low-level feature๋“ค๋งŒ ์‚ฌ์šฉํ–ˆ๋‹ค. ์ด์ƒ์ ์œผ๋กœ ๋ณด๋ฉด, Content image๋กœ๋ถ€ํ„ฐ ์ปจํ…์ธ ๋ฅผ ์ถ”์ถœํ•œ ๋‹ค์Œ Style image์˜ style์— ๋ Œ๋”๋งํ•˜๋Š” Texture transfer ์ ˆ์ฐจ๊ฐ€ ์ข‹์•„๋ณด์ธ๋‹ค.
  • 15. 15 Introduction ๊ทธ๋Ÿฌ๋ฏ€๋กœ ์ด๋ฏธ์ง€์˜ representations ์ฐพ๋Š” ๊ฒƒ์ด ์šฐ์„ ์ด ๋˜์–ด์•ผํ•œ๋‹ค. semantic image content์˜ variation์ด๋ž‘ ๊ทธ๋ฆผ์— ๋‚˜ํƒ€๋‚ผ ์Šคํƒ€์ผ์— ๋Œ€ํ•ด์„œ ๊ฐ๊ฐ ๋…๋ฆฝ์ ์œผ๋กœ ๋ชจ๋ธ๋งํ•˜๋Š” ๊ฒƒ์ด ์ด๋ฏธ์ง€์˜ representation์ด๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Ÿฌํ•œ factorized representations๋Š” ์ด๋ฏธ ์ด์ „์— ์ด๋ฃจ์–ด์ง„ ์  ์žˆ๋‹ค. ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ œํ•œ๋œ ์ด๋ฏธ์ง€์—์„œ ์ˆ˜ํ–‰๋˜์—ˆ๋‹ค. ๋‹ค๋ฅธ ์กฐ๋ช… ์•„๋ž˜์˜ ์–ผ๊ตด๋“ค, ๋‹ค๋ฅธ ํฐํŠธ ์Šคํƒ€์ผ์˜ ์ฒ ์ž๋“ค, ์†์œผ๋กœ ์“ฐ์—ฌ์ง„ ์ˆซ์ž๋“ค, ๋ฒˆ์ง€์ˆ˜ ๋“ฑ์ด ๊ทธ ์˜ˆ์‹œ๋‹ค.
  • 16. 16 Introduction ์Šคํƒ€์ผ๋กœ๋ถ€ํ„ฐ ์ปจํ…์ธ  ๋ถ„๋ฆฌ๋Š” ์—ฌ์ „ํžˆ ์–ด๋ ค์šด ๋ฌธ์ œ๋‹ค. ํ•˜์ง€๋งŒ Deep CNN์˜ ๋ฐœ์ „์€ ๊ฐ•๋ ฅํ•œ ์ปดํ“จํ„ฐ ๋น„์ „ ์‹œ์Šคํ…œ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ด์คฌ๋‹ค. ๋” ๊ตฌ์ฒด์ ์œผ๋กœ ๋งํ•˜์ž๋ฉด, Natural images์˜ high- level ์˜๋ฏธ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ฒƒ์„ ํ•™์Šต์‹œํ‚จ ์ปดํ“จํ„ฐ ๋น„์ „ ์‹œ์Šคํ…œ์ด ๊ฐ€๋Šฅํ•ด ์กŒ๋‹ค. ์ถฉ๋ถ„ํ•œ ์–‘์˜ ๋ ˆ์ด๋ธ”๋œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต์‹œํ‚จ CNN์€ ๋ฐ์ดํ„ฐ์…‹๋“ค ์ „๋ฐ˜์ ์œผ๋กœ ์ผ๋ฐ˜ํ™”๋˜์–ด ์žˆ๋Š” feature representations๋“ค์—์„œ high-level ์ด๋ฏธ์ง€ ์ปจํ…์ธ ๋ฅผ ์ถ”์ถœํ•˜๋Š” ๊ฒƒ์„ ํ•™์Šตํ•œ๋‹ค.
  • 17. 17 Introduction ์ด ๋…ผ๋ฌธ์—์„œ๋Š” deep CNN์˜๋กœ ํ•™์Šต๋œ generic feature representations๊ฐ€ natural images๋“ค์˜ ์ปจํ…์ธ ์™€ ์Šคํƒ€์ผ์„ ๋…๋ฆฝ์ ์œผ๋กœ ์กฐ์ž‘ํ•˜๊ณ  ๋…๋ฆฝ์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์— ์‚ฌ์šฉ๋˜๋Š”์ง€ ๋ณด์—ฌ์ค€๋‹ค. ๋ชจ๋ธ์ด ๋”ฅํ•˜๋‹ค๋ณด๋‹ˆ ์‹ฑ๊ธ€ ๋‰ด๋Ÿด๋„ท์—์„œ ์žˆ์—ˆ๋˜ ์˜ตํ‹ฐ๋งˆ์ด์ œ์ด์…˜ ๋ฌธ์ œ๋ฅผ ์ค„์ผ ์ˆ˜ ์žˆ๋‹ค. ์ƒˆ๋กœ์šด ์ด๋ฏธ์ง€๋Š” ์–ด๋–ป๊ฒŒ ๋งŒ๋“ค์–ด์ง€๋ƒ ํ•˜๋ฉด ์ƒ˜ํ”Œ์ด๋ฏธ์ง€์˜ feature representations์™€ ์–ด๋– ํ•œ pre- image๋ฅผ ๋งค์นญ ์‹œํ‚ค๋ฉด์„œ pre-image์˜ ๋ชจ์Šต์„ ์ฐพ์•„๋‚˜๊ฐ€๋Š” ๋ฐฉ๋ฒ•์ด๋‹ค. ์‹ค์ œ๋กœ, ์šฐ๋ฆฌ์˜ ์Šคํƒ€์ผ ํŠธ๋žœ์Šคํผ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ โ€˜image representationsโ€™์˜ ๋ฐ˜๋Œ€ ๋ฐฉ๋ฒ•์„ ์“ฐ๋Š” CNN์ด ๋ฒ ์ด์Šค์ธ model์„ ์“ด๋‹ค.
  • 18. 18 Deep image representations Pre-trained 19 layer VGG network ์‚ฌ์šฉ ์ถœ์ฒ˜: https://www.youtube.com/watch?v=fIW8fI2Xb_k
  • 19. 19 Deep image representations ์ถœ์ฒ˜: https://www.youtube.com/watch?v=fIW8fI2Xb_k
  • 20. 20 ๋’ค๋กœ ๊ฐˆ์ˆ˜๋ก ๋””ํ…Œ์ผ ํ”ฝ์…€ ์ •๋ณด๊ฐ€ ๋กœ์Šค๋œ๋‹ค. ์„œ๋กœ ๋‹ค๋ฅธ ๋ ˆ์ด์–ด๋“ค๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ถ„์„ํ•œ๋‹ค. conv1 / conv1,conv2 / conv1, conv2, conv3 / conv1,conv2,conv3,con4 / conv1,conv2,conv3, conv4, conv5 e: ๋งˆ์นจ๋‚ด global arrangement ๋Š” ์‚ฌ๋ผ์ ธ๋„ ์ฃผ์–ด์ง„ ์ด๋ฏธ์ง€์˜ ์Šคํƒ€์ผ์ด ์ž˜ ์ƒ์„ฑ๋œ๋‹ค.
  • 23. 23 Content representation โƒ—" : generated image โƒ—# : original image $% : original image์˜ feature representation in layer & '% : generated image์˜ feature representation in layer & ์œ„์˜ ๋กœ์Šค ํ•จ์ˆ˜๋ฅผ ํ†ตํ•ด gradient descen๋ฅผ ํ†ตํ•ด โƒ—"๊ฐ€ โƒ—#๊ฐ€ ๋˜๊ฒŒ๋” ํ•œ๋‹ค.
  • 24. 24 Style representation Layer ๋“ค ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ถ„์„ํ•˜๋Š” ํ‹€์ธ ๊ทธ๋žจ ๋งคํŠธ๋ฆญ์Šค๋ฅผ ์‚ฌ์šฉ ์ถœ์ฒ˜: ๋ณด์ฐฌ๋‹˜์˜ ์Šฌ๋ผ์ด๋“œ(CycleGAN) Gram Matrix ํ™์ •๋ชจ ๊ต์ˆ˜๋‹˜ ์œ ํˆฌ๋ธŒ ๋ฐ”๋กœ๊ฐ€๊ธฐ (ํด๋ฆญ!)
  • 28. 28 Trade-off between content and style matching
  • 29. 29 Effect of different layers of the Convolutional Neural Network
  • 33. 33 ๊ฝค ์ž˜ํ•œ ๊ฒƒ ๊ฐ™์ง€๋งŒ ๊ทธ๋ž˜๋„ ํ•œ๊ณ„์ ์ด ์žˆ๋‹ค. Discussion
  • 34. 34 LIMITATION 1 Resolution of the synthesized images. The speed of the synthesis procedure depends heavily on image resolution. Discussion
  • 35. 35 LIMITATION 2 Noise. While this is less of an is- sue in the artistic style transfer, the problem becomes more apparent when both, content and style images, are photographs. Discussion
  • 36. 36 LIMITATION 2Discussion ์ œ๊ฐ€ ์ง์ ‘ ํ•œ๋ฒˆ ํ…Œ์ŠคํŠธ ํ•ด๋ณด๊ฒ ์๋‹ˆ๋‹ค.. ์‚ฝ์งˆ ํ•  ์ค€๋น„๊ฐ€ ๋˜์–ด์žˆ๋Š” ๋ฌธ์ˆ˜์˜์”จ ์ด๋ฏธ ์‚ฝ์งˆ ์ค‘์ธ ์•„์ €์”จ
  • 37. 37 LIMITATION 2Discussion ์‚ฝ์งˆ ํ•  ์ค€๋น„๊ฐ€ ๋˜์–ด์žˆ๋Š” ๋ฌธ์ˆ˜์˜์”จ ์ด๋ฏธ ์‚ฝ์งˆ ์ค‘์ธ ์•„์ €์”จ LIMITATION 1์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ๋น ๋ฅธ ํ•™์Šต์„ ์œ„ํ•ด ๋””๋ฉ˜์…˜์€ 128*128๋กœ ๋ฐ”๊ฟจ์Šต๋‹ˆ๋‹ค...
  • 38. 38 LIMITATION 2Discussion ๋ฉ”๋งˆ๋ฅธ ๋•…์— ๊ฐ•๋ฌผ์ด ํ๋ฅด๊ฒŒ ๋˜๊ณ  ํ•˜๋Š˜์—๋Š” ํ’€์ด ์ž๋ผ๋‚˜๋‹ˆ
  • 39. 39 LIMITATION 2Discussion ์›๋ž˜๋ถ€ํ„ฐ ํ™”์งˆ์ด ์•ˆ์ข‹์•„์„œ ๋…ธ์ด์ฆˆ๊ฐ€ ์ƒ๊ธด์ง€ ์ž˜ ๋ชจ๋ฅด๊ฒ ์Šตโ€ฆ
  • 40. 40 LIMITATION 2Discussion ์˜ค๋ฆฌ์ง€๋‚  ์ด๋ฏธ์ง€๋ž‘ ๋น„๊ตํ•ด๋ณด๋ฉด ๋…ธ์ด์ฆˆ๊ฐ€ ์ข€ ์ƒ๊ธด ๊ฒƒ ๊ฐ™๊ธฐ๋„โ€ฆ
  • 41. 41 LIMITATION 2 However, the noise is very characteristic and appears to resemble the filters of units in the network. Discussion
  • 42. 42 LIMITATION 3 The separation of image content from style is not necessarily a well defined problem. This is mostly because it is not clear what exactly defines the style of an image. It might be the brush strokes in a painting, the colour map, certain dominant forms and shapes, but also the composition of a scene and the choice of the subject of the image and probably it is a mixture of all of them and many more. Discussion
  • 43. 43 LIMITATION 3Discussion ์˜ˆ๋ฅผ ๋“ค๋ฉด, ๋‚ด๊ฐ€ ๋ˆˆ์น ์™์‹ฑ์„ ํ•˜๊ณ  ์‹ถ์–ด์„œ ๋ชจ๋‚˜๋ฆฌ์ž ์Šคํƒ€์ผ์„ ๊ฐ€์ ธ๋‹ค๊ฐ€ ์“ฐ๊ฒ ๋‹ค๊ณ  ํ•ด์„œ ๋‚ด ๋ˆˆ์น์ด ์‚ฌ๋ผ์ง€๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋‹ค.. ๋‚ด ์ •๋ฉด์˜ ๋ชจ์Šต์ด ๋งˆ์Œ์— ์•ˆ ๋“ค์–ด์„œ ์ธก๋ฉด์„ ๊ธฐ๋Œ€ํ•˜๊ณ  ์ €๋Ÿฐ ๋‚จ์ž ์‚ฌ์ง„ ์Šคํƒ€์ผ์„ ๊ฐ€์ ธ๋‹ค ์“ด๋‹ค๊ณ  ํ•ด์„œ ๋‚ด๊ฐ€ ์–ผ๊ตด์„ ๋Œ๋ฆฌ๊ณ  ์ฐ์€ ์‚ฌ์ง„์ด ํ•ฉ์„ฑ๋˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋‹ค.. ์‚ฌ๊ธฐ ์‚ฌ์ง„ ๋“ฑ์žฅ ใ…Ž
  • 44. 44 LIMITATION 3 In our work we consider style transfer to be successful if the generated image โ€˜looks likeโ€™ the style image but shows the objects and scenery of the content image. ํ˜„์‹ค๊ณผ ํƒ€ํ˜‘ํ•  ์‹œ๊ฐ„โ€ฆ Discussion
  • 45. 45 Nevertheless, ์ƒ๋ฌผํ•™์  ์‹œ๊ฐ ์ฒด๊ณ„๊ฐ€ ์ˆ˜ํ–‰ํ•˜๋Š” ์ผ๋“ค ์ค‘ ํ•˜๋‚˜๋ฅผ ํ•ด๋‚ผ ์ˆ˜ ์žˆ๋„๋ก ํ›ˆ๋ จ๋œ Neural Network๊ฐ€ ์–ด๋Š ์ •๋„๋Š” style๋กœ๋ถ€ํ„ฐ contents๋ฅผ ์ž๋™์œผ๋กœ ๋ถ„๋ฆฌํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ด ๊ต‰์žฅํžˆ fascinatingํ•˜๋‹ค. Discussion
  • 46. 46 ๋งˆ๋ฌด๋ฆฌ ์ตœ์ ํ™”๋œ ANN์˜ ์„ฑ๋Šฅ๊ณผ ์‚ฌ๋žŒ์˜ ๋น„์ „์ด ์„œ๋กœ ์ƒ๋‹นํžˆ ๋‹ฎ์•˜๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•ด๋ณด๋ฉด ์‚ฌ๋žŒ๋„ ์—ญ์‹œ contents ๋ฅผ style๋กœ๋ถ€ํ„ฐ ์ถ”๋ก ํ•ด๋‚ด๋Š” ๋Šฅ๋ ฅ์ด ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋ ‡๊ธฐ ๋•Œ๋ฌธ์— ์˜ˆ์ˆ ์„ ์ฐฝ์กฐํ•˜๊ณ  ์ฆ๊ธธ ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์ด๋‹ค. ์ฆ‰, ์šฐ๋ฆฌ์˜ ์‹œ๊ฐ ์ฒด๊ณ„์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์— ๋Œ€ํ•œ ํ™•์‹คํ•œ ํŠน์„ฑ์„ ์œ ์ถ”ํ•  ์ˆ˜ ์žˆ๋‹ค. ์•„๋ฆ„๋‹ค์šด ๋งˆ๋ฌด๋ฆฌ Discussion