/prog/ - a touhou face recognizer

Name: Anonymous 2012-07-26 8:28

how difficult would it be to create a pipeline that can detect and recognize touhou faces in images?

I was thinking of building something very simple like this using opencv:
1. face detector that returns bounding boxes around possible faces. copying code from one of them generic face detector will probably do.

2. extract sift features from the cropped images in the bounding boxes. again copying code will be the easiest way

3. use a svm trained on 250+ images for each touhou downloaded from danbooru or something to classify the sift features from the cropped images

Thats all I can think of for now. Any suggestions for improving this very basic pipeline would be helpful.

Name: Anonymous 2012-07-26 8:29

Surely the hats would be easier to recognise.

Name: 2 2012-07-26 8:34

Even then, difficulties would arise due to the variety of artist styles:
http://safebooru.org//images/733/234e4c270554e50e3c221005eaa03e0eae04190a.jpg
http://safebooru.org//images/860/07b788c2b642c092e4cb2251652d1199fe4ac531.jpg
http://safebooru.org//images/870/992f49d69fd8e75b54ca71cf382cd9594ad3ba8a.jpg
http://safebooru.org//images/870/98e189ffdcaba638b92c44ac680717333a6dcc3b.png

Name: Anonymous 2012-07-26 8:36

op here
not every touhou has a hat, also we cant be certain they will be wearing the hat in the picture
also since each touhou has a hair/eye colors, maybe I can use histograms as features too.

Name: Anonymous 2012-07-26 8:41

>>2
yeah your right. In fact, if not the heart on her head, and her hair color, even I cant really tell which Touhou she is just by looking at her face

Name: Anonymous 2012-07-26 9:22

>>4
Every Touhou has a head ornament, NO EXCEPTIONS

Name: Anonymous 2012-07-26 10:02

Just have a script query http://iqdb.org/ for booru results then extract Touhou names from the image tags.

Name: >>7 2012-07-26 10:42

Like this:

require 'nokogiri'

require 'net/http'

require 'uri'

require 'open-uri'



touhous = []



doc = Nokogiri::HTML(open('http://danbooru.donmai.us/wiki/show?title=List_of_Touhou_Characters').read)

found_reimu = false



doc.css("a").each do |i|

  next unless i["href"] =~ /wiki.show/

  found_reimu ||= i["href"] =~ /reimu/

  next unless found_reimu

  touhous << URI.unescape(i["href"].split('=')[1])

end



image = ARGV[0]



response = Net::HTTP.post_form URI('http://iqdb.org/'), { "url" => image }

doc = Nokogiri::HTML(response.body)

tags = doc.css("img")[1]["alt"].split[3..-1]



raise "Image does not contain touhous!" unless tags.include? "touhou"



puts (tags & touhous)

Name: >>7 2012-07-26 10:48

If you're a sucker for pretty output, change that last line to:

puts (tags & touhous).map { |name| name.tr('_', ' ').split.map(&:capitalize).join(' ') }

Name: Anonymous 2012-07-26 11:41

>>8
Disgusting. Absolutely horrible.

Name: VIPPER 2012-07-26 11:51

>>9
'_'

Name: Anonymous 2012-07-26 12:31

>>10
What? Might be slow as balls, but it certainly is better than training a PC to recognize hats.

Name: >>7 2012-07-26 14:57

>>12
It could be a lot faster if I hardcode the Touhou list, but the iqdb POST request is still half of it. (And Ruby 1.9 is none of it, if that's what you mean.)

Name: >>7 2012-07-26 15:28

There.

require "nokogiri"

require "net/http"

require "uri"

require "open-uri"



response = Net::HTTP.post_form URI("http://iqdb.org/"), { url: ARGV[0] }

raise "image not found on iqdb" if response.body !~ /Best match/



# Extract tags from response

doc = Nokogiri::HTML(response.body)

tags = doc.css("img")[1]["alt"].split[3..-1]



# Look for touhous

touhous = DATA.read.split("\n")



(touhous & tags).each do |name|

  puts name.tr('_', ' ').split.map(&:capitalize).join(' ')

end



__END__

hakurei_reimu

hakurei_reimu_(classic)

kirisame_marisa

kirisame_marisa_(classic)

kirisame_marisa_(seihou)

genjii

shingyouku

yuugenmagan

elis_(touhou)

sariel

mima

kikuri_(touhou)

konngara

rika_(touhou)

meira

the_5_magic_stones

evil_eye_sigma

noroiko

rengeteki

ellen

kotohime

kana_anaberal

asakura_rikako

kitashirakawa_chiyuri

okazaki_yumemi

ruukoto

mimi-chan

orange_(touhou)

kurumi_(touhou)

elly

kazami_yuuka

kazami_yuuka_(classic)

mugetsu

gengetsu

sara_(touhou)

luize

yuki_(touhou)

mai_(touhou)

yumeko

shinki

alice_margatroid_(young)

rumia

ex-rumia

daiyousei

cirno

advent_cirno

cirno-nee

achi_cirno

ikazu_cirno

mizu_cirno

hong_meiling

ex-meiling

hong_meiling_(panda)

koakuma

patchouli_knowledge

izayoi_sakuya

inu_sakuya

remilia_scarlet

flandre_scarlet

letty_whiterock

chen

chen_(cat)

alice_margatroid

shanghai_doll

hourai_doll

lily_white

lunasa_prismriver

merlin_prismriver

lyrica_prismriver

konpaku_youmu

myon

saigyouji_yuyuko

saigyouji_yuyuko_(living)

yakumo_ran

yakumo_ran_(fox)

yakumo_yukari

saigyou_ayakashi

usami_renko

maribel_hearn

morichika_rinnosuke

mannosuke

tokiko_(touhou)

ibuki_suika

ibuki_suika_(watermelon)

wriggle_nightbug

princess_wriggle

mystia_lorelei

mystia_lorelei_(bird)

gothtia

kamishirasawa_keine

ex-keine

sukusuku_hakutaku

inaba_tewi

inaba_tewi_(bunny)

reisen_udongein_inaba

reisen_udongein_inaba_(bunny)

yagokoro_eirin

yagokoro

houraisan_kaguya

fujiwara_no_mokou

fujiwara_no_mokou_(young)

fujiwara_no_mokou_(phoenix)

luna_child

star_sapphire

sunny_milk

medicine_melancholy

su-san

onozuka_komachi

shameimaru_aya

shameimaru_aya_(crow)

shikieiki_yamaxanadu

lily_black

sunflower_fairy

hieda_no_akyuu

reisen

watatsuki_no_yorihime

watatsuki_no_toyohime

aki_shizuha

aki_minoriko

kagiyama_hina

kawashiro_nitori

kawashiro_nitori_(kappa)

inubashiri_momiji

inubashiri_momiji_(wolf)

mofuji

kochiya_sanae

yasaka_kanako

moriya_suwako

nagae_iku

hinanawi_tenshi

kisume

kurodani_yamame

mizuhashi_parsee

hoshiguma_yuugi

komeiji_satori

kaenbyou_rin

kaenbyou_rin_(cat)

zombie_fairy

reiuji_utsuho

reiuji_utsuho_(bird)

komeiji_koishi

nazrin

nazrin_(mouse)

tatara_kogasa

kumoi_ichirin

unzan

murasa_minamitsu

toramaru_shou

toramaru_shou_(tiger)

hijiri_byakuren

houjuu_nue

hisou_tensoku

namazu

goliath_doll

himekaidou_hatate

hata-tan

ibaraki_kasen

kasodani_kyouko

miyako_yoshika

kaku_seiga

soga_no_tojiko

mononobe_no_futo

toyosatomimi_no_miko

futatsuiwa_mamizou

futatsuiwa_mamizou_(tanuki)

hijiri_myouren

konpaku_youki

layla_prismriver

lord_tenma

satsuki_rin

vivit

Name: Anonymous 2012-07-26 16:47

Lush or Lush2 has native camera sampling and contrast detection features, as well as Lush is designed for complication scanning nets, mostly for AI navigation purposes, but the principles still apply.

Name: Anonymous 2012-07-26 17:06

>>8
image = ARGV[0]
wut? image is the command name?

Name: Anonymous 2012-07-26 17:15

>>16
ARGV contains arguments, NO EXCEPTIONS.

Name: Anonymous 2012-07-26 17:58

>>17
except argv[0] is the program/command name

Name: Anonymous 2012-07-26 18:06

>>18
Not in Ruby.

Name: Anonymous 2012-07-27 17:18

>>19
Wow, that's fucking stupid then.

Name: Anonymous 2013-09-01 23:09

There is a transfinite sequence of cardinal numbers:
0, 1, 2, 3, ..., n; ℵ₀, ℵ₀, ℵ₁, ... ℵ_α, ...

a touhou face recognizer

1 Name: Anonymous 2012-07-26 8:28

2 Name: Anonymous 2012-07-26 8:29

3 Name: 2 2012-07-26 8:34

4 Name: Anonymous 2012-07-26 8:36

5 Name: Anonymous 2012-07-26 8:41

6 Name: Anonymous 2012-07-26 9:22

7 Name: Anonymous 2012-07-26 10:02

8 Name: >>7 2012-07-26 10:42

9 Name: >>7 2012-07-26 10:48

10 Name: Anonymous 2012-07-26 11:41

11 Name: VIPPER 2012-07-26 11:51

12 Name: Anonymous 2012-07-26 12:31

13 Name: >>7 2012-07-26 14:57

14 Name: >>7 2012-07-26 15:28

15 Name: Anonymous 2012-07-26 16:47

16 Name: Anonymous 2012-07-26 17:06

17 Name: Anonymous 2012-07-26 17:15

18 Name: Anonymous 2012-07-26 17:58

19 Name: Anonymous 2012-07-26 18:06

20 Name: Anonymous 2012-07-27 17:18

21 Name: Anonymous 2013-09-01 23:09