๐Ÿ“‘ Paper/๐Ÿ–ผ๏ธ Computer Vision

Affective Image Classification using Features Inspired by Psychology and Art Theory

Gamddalki 2024. 2. 13. 20:59

Affective Image Classification using Features Inspired by Psychology and Art Theory (2010)

 

ABSTRACT

  • ์ด๋ฏธ์ง€๋Š” ๊ฐ์ •์ ์ธ ์ˆ˜์ค€์—์„œ ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Œ
    • ๊ทธ๋Ÿฌ๋‚˜ ์ด๋ฏธ์ง€๋ฅผ ๋ณด๋Š” ์‚ฌ๋žŒ์—๊ฒŒ ๋ฐœ์ƒํ•˜๋Š” ๊ฐ์ •์€ ๋งค์šฐ ์ฃผ๊ด€์ ์ด๊ธฐ ๋•Œ๋ฌธ์— ์ธ๋ฑ์Šค๊ฐ€ ํ˜•์„ฑ๋˜๋Š” ๊ฒฝ์šฐ๋Š” ๊ฑฐ์˜ ์—†์Œ
      • BUT ๊ฐ์ •์ ์ธ ๋‚ด์šฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ด๋ฏธ์ง€๋ฅผ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋‹ค๋ฉด ๋„์›€์ด ๋  ์ˆ˜ ์žˆ๋Š” ์ƒํ™ฉ์ด ์žˆ์Œ
  • ํ•ด๋‹น ๋…ผ๋ฌธ์€ ์ด๋ฏธ์ง€์˜ ๊ฐ์ •์ ์ธ ๋‚ด์šฉ์„ ๋‚˜ํƒ€๋‚ด๋Š” low level feature๋ฅผ ์ถ”์ถœํ•˜๊ณ  ๊ฒฐํ•ฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์กฐ์‚ฌํ•˜๊ณ  ๊ฐœ๋ฐœํ•˜์—ฌ ์ด๋ฏธ์ง€ ๊ฐ์ • ๋ถ„๋ฅ˜์— ์‚ฌ์šฉํ•จ
    • ์‹ฌ๋ฆฌํ•™๊ณผ ์˜ˆ์ˆ  ์ด๋ก ์˜ ์ด๋ก ์ ์ด๊ณ  ๊ฒฝํ—˜์ ์ธ ๊ฐœ๋…์„ ํ™œ์šฉํ•˜์—ฌ ๊ฐ์ •์ ์ธ ํ‘œํ˜„์œผ๋กœ ์˜ˆ์ˆ  ์ž‘ํ’ˆ ์˜์—ญ์— ๊ณ ์œ ํ•œ ์ด๋ฏธ์ง€ ํŠน์ง•์„ ์ถ”์ถœํ•จ
    • → IAPS์—์„œ SOTA์™€ ๋น„๊ตํ•˜์—ฌ ํ–ฅ์ƒ๋œ ๋ถ„๋ฅ˜ ๊ฒฐ๊ณผ๋ฅผ ์–ป์Œ

 

SYSTEM FLOW OF THE FRAMEWORK

1. image ๋ฐ์ดํ„ฐ ์…‹ 3์ข… ์ด์šฉ

  • IAPS(International Affective Picture System) (394๊ฐœ)
  • Artistic Photography ๊ณต์œ  ์‚ฌ์ดํŠธ ํฌ๋กค๋ง (807๊ฐœ)
    • using the emotion categories as search terms in the art sharing site
      • These photos are taken by people who attempt to evoke a certain emotion in the viewer
      • ์ž‘๊ฐ€๊ฐ€ ํ‘œ์‹œํ•œ ์ƒ‰์ƒ๊ณผ ์งˆ๊ฐ์„ ์˜์‹์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๋ฉด ๋ถ„๋ฅ˜๊ฐ€ ๊ฐœ์„ ๋˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ
  • Abstract Painting (228๊ฐœ)
    • ํŠน์ • ๋ฌผ์ฒด, ์‹ฌ๋ณผ ๋“ฑ์ด ์œ ๋ฐœํ•˜๋Š” ๊ฐ์ •์„ ๋ฐฐ์ œํ•˜๊ณ  ๋งฅ๋ฝ์ ์ธ ๋‚ด์šฉ์ด ์—†๋Š” ์ƒํƒœ์—์„œ ์ด๋ฏธ์ง€์˜ ํŠน์ง•๋งŒ์„ ๋ณด๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ

2. Preprocessing ์ง„ํ–‰ ํ›„ segmentation

⇒ segmentation result + original image: input to the feature extraction process

Preprocessing

  • resize image into 200,000 pixels
  • crop borders
  • conversion from RGB to HSV
    • intuitive definition of colors by defining a well separated Hue (H), Saturation (S) and Brightness (Y) channel

Segmentation

  • Waterfall segmentation ์ด์šฉ
    • takes both color and spatial information into account
    • ⇒ results in an image separated into contiguous regions

3. FEATURE extraction: get the feature vector (low level feature ์ด์šฉ)

FEATURES

  • based on the experimentally-determined relation between ์ƒ‰์˜ ์ฑ„๋„, ๋ฐ๊ธฐ, ๊ทธ๋ฆฌ๊ณ  ๊ฐ์ •์˜ dimensions [26], as well as features based on relations between ์ƒ‰์˜ ํ˜ผํ•ฉ๊ณผ induced emotional effects from art theory [15].
    • We complement these features by a selection of features, some of which are shown to be of use in similar image retrieval [24] and classification tasks [7, 32]

1) Color

  • low-level color features๋ฅผ ๊ฐ์ •๊ณผ ๋งคํ•‘ํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์–ด๋ ค์šด ์ผ์ž„
    • ์ƒ‰์˜ ์‚ฌ์šฉ์— ๋Œ€ํ•œ ์ด๋ก ๊ณผ ์ธ์ง€ ๋ชจ๋ธ์„ ๊ณ ๋ คํ•ด์•ผ ํ•˜๋ฉฐ ๋ฌธํ™”์ , ์ธ๋ฅ˜ํ•™์  ๋ฐฐ๊ฒฝ์„ ํฌํ•จํ•ด์•ผ ํ•จ [15, 5]
      • ๋‹ค๋ฅธ ๋ฌธํ™”๋‚˜ ๋ฐฐ๊ฒฝ์„ ๊ฐ€์ง„ ์‚ฌ๋žŒ๋“ค์€ ๋™์ผํ•œ ์ƒ‰์ƒ ํŒจํ„ด์„ ์ƒ๋‹นํžˆ ๋‹ค๋ฅด๊ฒŒ ์ธ์‹ํ•˜๊ณ  ํ•ด์„ํ•  ์ˆ˜ ์žˆ์Œ (= ์šฐ๋ฆฌ์˜ ํ•œ๊ณ„์  ๋˜ํ•œ ๋  ์ˆ˜ ์žˆ์„ ๋“ฏ)
    • ์ƒ‰์ฑ„์™€ ์ƒ‰์ฑ„ ์กฐํ•ฉ์˜ emotional impact๋Š” ์˜ˆ์ˆ ๊ฐ€[15], ์‹ฌ๋ฆฌํ•™[26], ์ƒ‰์ฑ„ ๊ณผํ•™์ž[23] ๋ฐ ๋งˆ์ผ€ํŒ… ์—์ด์ „ํŠธ์˜ ๊ด€์ ์—์„œ ์กฐ์‚ฌ๋˜์—ˆ์Œ
  • ์ƒ‰์˜ ์ธก์ •์„ ์œ„ํ•ด implemented ๋œ features:
    • Saturation and Brightness statistics (์ฑ„๋„์™€ ๋ฐ๊ธฐ)
      • ์ง์ ‘์ ์ธ ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ์Œ on pleasure, arousal and dominance, the three axes of the emotion space according to the dimensional approach to emotions [22].
      • formulae determined by Valdez and Mehrabian from their ์‹ฌ๋ฆฌํ•™์  ์‹คํ—˜ [26].
        • The experiments were conducted in a controlled environment, where 250 people were shown series of single color patches and rated them on a standardized emotional scale (Pleasure-Arousal-Dominance) to describe how they feel about the color.
        • The resulting equations:
          • Pleasure = 0.69 Y +0.22 S (1)
          • Arousal = −0.31 Y +0.60 S (2)
          • Dominance = 0.76 Y +0.32 S (3)
    • Hue statistics (์ƒ‰์ƒ)
      • tone of the image๋ฅผ ๊ฒฐ์ •
      • ๊ทธ๋Ÿฌ๋‚˜ hue๋Š” circular way (in degrees)๋กœ ์ธก์ •๋จ
        • vector-based, circular statistics [20] must be used to compute measurements like mean, hue spread, etc.
    • Colorfulness
      • measured by Earth Mover’s Distance (EMD) between the histogram of an image and the histogram having a uniform color distribution, according to an algorithm suggested by Datta [7].
    • Color Names
      • each color has a special meaning and is used in certain ways by artists
        • 11๊ฐ€์ง€ ๊ธฐ๋ณธ ์ƒ‰์ƒ์— ๋Œ€ํ•ด ์‚ฌ์šฉ๋œ ํ”ฝ์…€ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•จ (black, blue, brown, green, gray, orange, pink, purple, red, white, yellow) are present on the image using the algorithm of van de Weijer et al. [27].
    • Itten contrasts [15]
      • a powerful concept in art theory
      • viewer์—๊ฒŒ ๊ฐ์ •์ ์ธ ํšจ๊ณผ๋ฅผ ์œ ๋„ํ•˜๊ณ  ์กฐํ™”๋กœ์šด ์ด๋ฏธ์ง€๋ฅผ ์–ป๊ธฐ ์œ„ํ•ด ์ƒ‰์„ ์กฐํ•ฉํ•˜๊ธฐ ์œ„ํ•œ ๊ณต์‹ํ™”๋œ ๊ฐœ๋…
      • hue, saturation and luminance(์ƒ‰์ƒ, ์ฑ„๋„, ํœ˜๋„)์— ๋”ฐ๋ผ ์ƒ‰์ƒ์„ characterizeํ•จ
        • 12๊ฐœ ์ƒ‰์ƒ์ด ๊ธฐ๋ณธ ์ปฌ๋Ÿฌ๋กœ identified
          • 5 ๋ ˆ๋ฒจ์˜ ํœ˜๋„์™€ 3 ๋ ˆ๋ฒจ์˜ ์ฑ„๋„๋กœ varied
          ⇒ 180๊ฐœ์˜ distinct colors, which have been organized into a spherical representation
        • 12๊ฐœ์˜ pure colors๊ฐ€ equatorial circle์— ๋”ฐ๋ผ ๋ฐฐ์—ด, luminance ํœ˜๋„๋Š” meridians์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋ฉฐ, radius๊ฐ€ ์ปค์ง์— ๋”ฐ๋ผ saturation์ด ์ฆ๊ฐ€
        • The center of the sphere is neutral gray, ๋Œ€์กฐ์ ์ธ ์ƒ‰์ƒ์€ ์„œ๋กœ ๋ฐ˜๋Œ€ํŽธ์— ์œ„์น˜ํ•จ
          • Warm colors lie opposite cold colors, dark colors opposite light colors, etc.
          • ์ฐธ๊ณ 

      • ์ด polar representation์„ ๋ฐ”ํƒ•์œผ๋กœ, Itten์€ 7๊ฐ€์ง€ ํƒ€์ž…์˜ contrast๋ฅผ identified
      • ๋˜ํ•œ formalized color combinations that look harmonious: the color accordances (์ƒ‰์˜ ์กฐํ™”)
      • ์ด๋ฅผ mathematical formulae๋กœ ๋‚˜ํƒ€๋ƒ„
        • partly based on the work in [6].
          1. transform the image into a simpler collection of colored patches
            1. waterfall segmentation์˜ ๊ฒฐ๊ณผ๋กœ ์–ป์€ region๊ณผ ๊ฐ๊ฐ์˜ average Hue, Saturation and Brightness ๊ณ„์‚ฐ ๊ฒฐ๊ณผ๋ฅผ ์ด์šฉ
            2. translate the region’s average values into the Itten color model [15]
            3. → ๊ฐ region์ด e.g. “dark”, with “low saturation” and “green hue”์™€ ๊ฐ™์ด ํ‘œํ˜„๋จ
          2. ๋ชจ๋ธ์˜ ์ฑ„๋„์™€ ๋ฐ๊ธฐ๋Š” fuzzy membership functions defined in [32]๋ฅผ ์ด์šฉํ•ด encoded๋จ
            1. ์ถ”๊ฐ€๋กœ ๊ฐ ์˜์—ญ(has Itten color and a size)์— ๋Œ€ํ•ด Itten contrasts๋ฅผ ๋ถ„์„ํ•จ
            2. ๋ช…์•”์˜ ๋Œ€์กฐ ์ธก์ •: ์ƒ๋Œ€์  ํฌ๊ธฐ์— ๋”ฐ๋ผ ๊ฐ€์ค‘์น˜๊ฐ€ ๋ถ€์—ฌ๋œ ๋ชจ๋“  ์˜์—ญ์˜ Brightness membership functions์— ๋Œ€ํ•œ ํ‘œ์ค€ ํŽธ์ฐจ ๋ฐ ์•„๋‚ ๋กœ๊ทธ ๋ฐฉ์‹์œผ๋กœ ํฌํ™”๋„์˜ ๋Œ€๋น„๋ฅผ ์ •์˜
            3. ์ƒ‰์ƒ์˜ ๋Œ€์กฐ ์ธก์ •: ๋ฒกํ„ฐ ๊ธฐ๋ฐ˜์˜ measurement of the hue spread
            4. complements์˜ ๋Œ€์กฐ ์ธก์ •: ๋ถ„ํ• ๋œ ์˜์—ญ ์‚ฌ์ด์˜ ์ƒ‰์ƒ ์ฐจ์ด๋ฅผ ๊ณ„์‚ฐ
              1. hue-wheel problem์„ ๊ณ ๋ คํ•ด์•ผ ํ•˜๋ฏ€๋กœ, hue difference measure (์ƒ‰์ฐจ ์ธก๋„, where hi is the representative (mean) hue of the region i)๋กœ d = min(|hi − hj |, 360 − |hi − hj |) ๋ฅผ ์‚ฌ์šฉํ•จ
                1. ๋งŒ์•ฝ contrast of complements๊ฐ€ ์žˆ๋‹ค๋ฉด ๊ฐ’์€ 180โ—ฆ์— ๊ฐ€๊นŒ์›€
            5. ์›œ์ฟจ์˜ ๋Œ€์กฐ ์ธก์ •: [6]์—์„œ ์ •์˜
              1. ๊ฐ ์˜์—ญ์—๋Š” ri ์˜์—ญ์˜ ์ถ”์œ„(t = 1), ์ค‘๋ฆฝ(t = 2) ๋ฐ ๋”ฐ๋œปํ•œ(t = 3) ์ •๋„๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” three membership functions wt๊ฐ€ ํ• ๋‹น
              2. warm-cold contrast between two regions is defined as:
              3. total amount of warm and cold area in the image ๋˜ํ•œ ๊ณ„์‚ฐํ•จ
            6. simultaneous contrast: absence of contrast of complements์ž„
              1. complementary contrast๊ฐ’์ด ๋‚ฎ์€ ๊ฒƒ๊ณผ ๊ฐ™์Œ
            7. contrast of extension: ์ˆ˜ํ•™์ ์œผ๋กœ ๊ณต์‹ํ™” ๋˜๊ธฐ ์–ด๋ ค์›Œ ๊ณ„์‚ฐํ•˜์ง€ ์•Š์Œ
            8. Harmony: ์‚ฌ๋žŒ์˜ ๋ˆˆ์— ์•ˆ์ •๊ฐ์„ ์ฃผ๋Š” ์ƒ‰์ƒ๊ณผ ํ†ค์˜ ์กฐํ•ฉ
              1. the combination of those colors whose mix is gray๋กœ ์ •์˜๋˜๋Š” ๊ฐ๊ด€์ ์ธ ๊ฐœ๋…์ž„
              2. spherical model์— ์ ์šฉ๋œ color accordances ๋Š” ๊ตฌ๋ฉด ์ƒ ์œ„์น˜๋ฅผ ์„ ์œผ๋กœ ์—ฐ๊ฒฐํ–ˆ์„ ๋•Œ ์ •๋‹ค๊ฐํ˜•์„ ์ƒ์„ฑํ•˜๋Š” ์ƒ‰์ƒ ์กฐํ•ฉ์ž„
              3. ๊ณ„์‚ฐ ๋ฐฉ๋ฒ•
                1. ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ 12๊ฐœ ์ฃผ์š” ์ƒ‰์ƒ์„ ๊ฐ€์ง„ hue histogram์„ ์ž‘์„ฑํ•จ (5% ๋ฏธ๋งŒ์ธ bin์€ ๋ฌด์‹œ)
                  1. ์ฃผ๋กœ 3–4๊ฐœ์˜ ์ฃผ์š” ์ƒ‰์ƒ์ด ๋ฐœ๊ฒฌ๋จ
                2. ์œ„ ์ฃผ์š” ์ƒ‰์ƒ์„ Itten color wheel๊ณผ mappingํ•˜๊ณ , ์ด๋“ค์˜ ์œ„์น˜๋ฅผ ์—ฐ๊ฒฐํ•ด ๋‹ค๊ฐํ˜• ์ƒ์„ฑ
                3. ์ƒ์„ฑ๋œ ๋‹ค๊ฐํ˜•์˜ ๋‚ด๊ฐ๊ณผ ๋™์ผํ•œ ์ˆ˜์˜ ์ •์ ์„ ๊ฐ€์ง„ ๊ฐ€์ƒ์˜ ์ •๋‹ค๊ฐํ˜•์˜ ๋‚ด๊ฐ์˜ ์ฐจ์ด๋กœ ์ธก์ •
  • Color features by Wang Wei-ning et al. [32]
    • ์ƒ‰์ƒ์˜ ๊ฐ์ •์  ์˜ํ–ฅ์„ ์ด๋ฏธ์ง€๋กœ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋„๋ก ์„ค๊ณ„๋œ ์ „๋ฌธ ํžˆ์Šคํ† ๊ทธ๋žจ

2) Texture

  • ์ „๋ฌธ ์‚ฌ์ง„์ž‘๊ฐ€๋“ค๊ณผ ์˜ˆ์ˆ ๊ฐ€๋“ค์€ ๋ณดํ†ต ์„ ๋ช…ํ•œ ์‚ฌ์ง„, ํ˜น์€ ์•„์›ƒํฌ์ปค์‹ฑ(the main object is sharp with a blurred background) ๋œ ์‚ฌ์ง„์„ ๋งŒ๋“ฆ
    • ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ๋Š” ์‚ฌ์ง„ ์† ๋ธ”๋Ÿฌ ์ฒ˜๋ฆฌ๊ฐ€ desired expression์„ achieveํ•˜๋Š” ๋ฐ์—๋„ ํšจ์œจ์ ์œผ๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ•จ
      • ์˜๋„์ ์œผ๋กœ blur๋œ ์ด๋ฏธ์ง€๋Š” fear๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” art photography images์—์„œ ์ž์ฃผ ํ™œ์šฉ๋˜์—ˆ์Œ
  • Wavelet-based features
    • to measure spatial smoothness/graininess in images
      • using the Daubechies wavelet transform [8]
    • [7]์—์„œ ์ œ์•ˆ๋œ three-level wavelet transform on all three color channels (Hue H, Saturation S and Brightness Y)๋ฅผ ์ด์šฉํ•จ
      • ์˜์ƒ์˜ ํ•œ ์ฑ„๋„์˜ wavelet transform์— ๋Œ€ํ•œ ์ˆ˜์ค€ i์˜ ๊ณ„์ˆ˜๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ํ‘œ์‹œํ•  ๋•Œ:
        • the wavelet features (where i = {1, 2, 3}) are:
           
          • ๋ชจ๋“  ๋ ˆ๋ฒจ i์™€ ๋ชจ๋“  ์ฑ„๋„(H, S, Y)์— ๋Œ€ํ•ด ๊ณ„์‚ฐ๋จ
            • ์ฆ‰, 9๊ฐœ์˜ wavelet features๋ฅผ ์–ป๊ฒŒ ๋จ
            • sum over all three levels๋ฅผ ๊ฐ ์ฑ„๋„๋งˆ๋‹ค ์ถ”๊ฐ€ํ•จ (์ด 3๊ฐœ)
  • Tamura [25]
    • affective image retrieval [33]์—์„œ ํšจ๊ณผ์ ์œผ๋กœ ์‚ฌ์šฉ๋˜์—ˆ์Œ
      • coarseness, contrast and directionality
    • ⇒ ๋”ฐ๋ผ์„œ Tamura texture features ์ค‘ ์•ž์—์„œ 3๊ฐœ๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ๋กœ ํ•จ:
  • Gray-Level Co-occurrence Matrix (GLCM) [13]
    • ์งˆ๊ฐ์„ ์ธก์ •ํ•˜๋Š” ๋˜ ํ•˜๋‚˜์˜ ๊ณ ์ „์ ์ธ ๋ฐฉ๋ฒ•์ž„
    • ⇒ compute contrast, correlation, energy, and homogeneity of an image

3) Composition

  • ์ด๋ฏธ์ง€ part ๊ฐ„์˜ ๊ด€๊ณ„๋Š” ๊ต‰์žฅํžˆ ๋ณต์žกํ•˜๋ฏ€๋กœ only few aspects of composition๋งŒ analyzeํ–ˆ์Œ
    • ๊ทธ๋Ÿฌ๋‚˜ ์ด ๋ถ„์•ผ๋Š” ์•ž์œผ๋กœ much improvement could be made
  • Level of Detail
    • detail์ด ๋งŽ์€ ์ด๋ฏธ์ง€๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฏธ๋‹ˆ๋ฉ€๋ฆฌ์ŠคํŠธ ๊ตฌ์„ฑ๊ณผ ๋‹ค๋ฅธ ์‹ฌ๋ฆฌ์  ํšจ๊ณผ๋ฅผ ๋ƒ„
    • ์ด๋ฏธ์ง€์˜ detail ์ˆ˜์ค€์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด predefined Alternating Sequential Filter size(ํ•„ํ„ฐ ํฌ๊ธฐ 3 ๋ฐ ํญํฌ ๊ณ„์ธต์˜ ์ˆ˜์ค€ 2)๋กœ waterfall segmentatione๋œ ์˜์—ญ์˜ ๊ฐœ์ˆ˜๋ฅผ countํ•จ
      • simple images: low, busy or cluttered images: high
  • Low Depth of Field (DOF)
    • ์ „๋ฌธ ์‚ฌ์ง„์ž‘๊ฐ€๊ฐ€ ๋ฐฐ๊ฒฝ์„ ๋ธ”๋Ÿฌ ์ฒ˜๋ฆฌํ•˜์—ฌ reducing the “busyness”ํ•˜๊ณ  attention of the observer๋ฅผ object of interest๋กœ ์ง‘์ค‘๋˜๊ฒŒ๋” ํ•˜๋Š” ๋ฐ์— ์‚ฌ์šฉ๋จ (์•„์›ƒํฌ์ปค์‹ฑ)
    • [7]์—์„œ ์ „์ฒด ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ์ด๋ฏธ์ง€์˜ ๋‚ด๋ถ€ ๋ถ€๋ถ„์˜ ๊ณ ์ฃผํŒŒ์ˆ˜(์‹ 4์˜ ํ‘œ๊ธฐ์— ์‚ฌ์šฉ๋œ ๋ ˆ๋ฒจ 3)์˜ ์›จ์ด๋ธ”๋ฆฟ ๊ณ„์ˆ˜์˜ ๋น„์œจ์„ ๊ณ„์‚ฐํ•˜์—ฌ low DOF ๋ฐ macro image๋ฅผ ๊ฒ€์ถœํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ์ œ์•ˆ๋จ
  • Dynamics
    • ์„ ํ–‰ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์„ ๋“ค์€ emotional effects [15, 2]๋ฅผ ์œ ๋ฐœํ•จ
      • Horizontal lines(์ˆ˜ํ‰์„ ): associated with a static horizon(์ •์ ์ธ ์ง€ํ‰์„ ) and communicate calmness, peacefulness and relaxation
      • vertical lines(์ˆ˜์ง์„ ): clear and direct and communicate dignity and eternity
      • slant lines(๊ธฐ์šธ์–ด์ง„ ์„ ): unstable and communicate dynamism
      • Lines with many different directions: chaos, confusion or action
      • ์„ ์ด ๊ธธ๊ณ  ๊ตต์œผ๋ฉฐ ์ง€๋ฐฐ์ ์ผ์ˆ˜๋ก induced psychological effect๊ฐ€ ๊ฐ•ํ•จ
    • Hough transform์„ ํ†ตํ•ด significant line slopes๋ฅผ detectํ•จ
      • ๋ฐœ๊ฒฌ๋œ ์„ ์€ ๊ธฐ์šธ๊ธฐ ๊ฐ๋„ θ์— ๋”ฐ๋ผ static (horizontal and vertical) or slant๋กœ ๋ถ„๋ฅ˜๋˜๊ณ , ๊ฐ๊ฐ์˜ ๊ธธ์ด๋กœ ๊ฐ€์ค‘์น˜๊ฐ€ ๋ถ€์—ฌ๋จ
        • (-15 โ—ฆ < θ < โ—ฆ < โ—ฆ>) ๋˜๋Š” (75 โ—ฆ < θ < 105 โ—ฆ)์ด๋ฉด ์ •์ , ๊ทธ๋ ‡์ง€ ์•Š์œผ๋ฉด ๊ธฐ์šธ๊ธฐ
      ⇒ ์ด๋ฏธ์ง€์—์„œ ์ •์  ์„ ๊ณผ ๋™์  ์„ ์˜ ๋น„์œจ์„ ์–ป์Œ
  • Rule of Thirds
    • main object๋Š” ๋‚ด๋ถ€ ์ง์‚ฌ๊ฐํ˜•์˜ ์•ˆ์ชฝ ํ˜น์€ ์ฃผ๋ณ€์— ๋†“์—ฌ ์žˆ์Œ
    • ⇒ ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ๋‚ด๋ถ€ ์ง์‚ฌ๊ฐํ˜•์— ๋Œ€ํ•œ color statistics๋ฅผ ์ธก์ •ํ•˜์˜€์Œ

4) Content

  • ์ด๋ฏธ์ง€์˜ semantic content๋Š” ์–ด๋–ค ๊ทธ๋ฆผ์˜ emotional influence์— ๊ฐ€์žฅ ํฐ ์˜ํ–ฅ์„ ๋ฏธ์นจ
  • induced two algorithms
    • Human Faces
      • strongly draw the attention of human observers
      • ํ‘œ์ •์€ distinguish the moods of a picture๋ฅผ ์œ„ํ•ด ๋งค์šฐ ์ค‘์š”ํ•˜๋‚˜, ์ด๋ฏธ์ง€์—์„œ ์‚ฌ๋žŒ์˜ ํ‘œ์ •์„ ์ธ์‹ํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์•„์ง ์™„๋ฒฝํ•˜์ง€ ์•Š์Œ
        • ๋”ฐ๋ผ์„œ (ํ•˜์œ„ํ˜ธํ™˜์œผ๋กœ) ์–ผ๊ตด ๊ฐ์ง€ ์•Œ๊ณ ๋ฆฌ์ฆ˜ by Viola and Jones [28]์„ ์‚ฌ์šฉํ•˜์˜€์Œ
        • ๋ฐœ๊ฒฌ๋œ ์–ผ๊ตด์˜ ์ˆ˜, ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ๊ฐ€์žฅ ํฐ ์–ผ๊ตด์˜ ์ƒ๋Œ€์ ์ธ ํฌ๊ธฐ๋ฅผ ์‚ฌ์šฉํ•จ
        • → ์ธ๋ฌผ์‚ฌ์ง„ ๊ตฌ๋ถ„, ๋‹จ์ฒด์‚ฌ์ง„๊ณผ ์ดˆ์ƒํ™” ๊ตฌ๋ถ„์ด ๊ฐ€๋Šฅํ•จ
    • Skin
      • ์Šคํ‚จ ์ปฌ๋Ÿฌ ํ”ฝ์…€์˜ ์–‘์œผ๋กœ “artistic nudes” ์ด๋ฏธ์ง€๋ฅผ ๊ตฌ๋ณ„ํ•จ
        • very specific emotional response๋ฅผ ๊ฐ–๋Š” ์ด๋ฏธ์ง€์ž„
      • [18]์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ด์šฉํ•จ
        • basic idea: ์ด๋ฏธ์ง€์—์„œ ํ”ผ๋ถ€์ƒ‰์„ ๋‚˜ํƒ€๋‚ด๋Š” ์ƒ‰์ƒ ์ŠคํŽ™ํŠธ๋Ÿผ์„ ์ฐพ๋Š” ๊ฒƒ
          • YCbCr color space is well suited to this task
            • ์‚ฌ์ „์— ์ •์˜๋œ static model (thresholds on the Cb and Cr channels)
              • many case์— ๋Œ€ํ•ด ํ”ผ๋ถ€์ƒ‰์„ ์ž˜ ๋‚˜ํƒ€๋ƒ„
        • [18]์—์„œ๋Š” face detection์„ ํฌํ•จํ•ด ์œ„ ๋ฐฉ๋ฒ•์„ ๊ฐœ์„ ํ•จ
          • face detection algorithm by Viola and Jones [28]๋ฅผ ์ด์šฉํ•ด ์–ผ๊ตด ๋จผ์ € ๊ฐ์ง€→ ๋งŒ์•ฝ ํ•˜๋‚˜ ์ด์ƒ์˜ ์–ผ๊ตด์ด ๋ฐœ๊ฒฌ๋œ๋‹ค๋ฉด skin color models are combined
          • → ๋งŒ์•ฝ ์–ผ๊ตด์ด ๋ฐœ๊ฒฌ๋œ๋‹ค๋ฉด the thresholds in the above model are altered to present a spectrum specific to the person found in the image
          • ์šฐ๋ฆฌ๋Š” ํ”ผ๋ถ€ ๋ฉด์  (i.e. the number of pixels in skin color)๊ณผ ๊ฐ์ง€๋œ ์–ผ๊ตด ํฌ๊ธฐ์— ๋Œ€ํ•œ proportion of the “skin area”๋ฅผ feature๋กœ ๊ณ„์‚ฐํ•˜์˜€์Œ

4. Classifier: Naive Bayes classifier ์ด์šฉ

 

RESULTS

  • best feature selection for each data set and each category, along with the dimensionality reduction method producing each feature set listed in the last row
    • IAPS
      • Anger ๋ฐ Contentment: Yanulevskaya et al. features๊ฐ€ ๊ฐ€์žฅ ์ ํ•ฉํ•œ ์„ฑ๋Šฅ์„ ์ œ๊ณต
      • IAPS ์„ธํŠธ์˜ ์นดํ…Œ๊ณ ๋ฆฌ๋Š” ์ฝ˜ํ…์ธ ์™€ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์ด ์žˆ์Œ
        • Fear ๋ฐ Disgust: snakes, insects, injuries๊ฐ€ ์ž์ฃผ ํ‘œ์‹œ
        • Amusement: ํ–‰๋ณตํ•œ ์‚ฌ๋žŒ์˜ ์ดˆ์ƒํ™”๊ฐ€ ํฌํ•จ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Œ
    • best feature set๋Š” ์นดํ…Œ๊ณ ๋ฆฌ์™€ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋ชจ๋‘์— ์˜์กดํ•จ
      • Amusement: ์‚ฌ๋žŒ์˜ ์–ผ๊ตด occurrence์™€ ํฌ๊ธฐ๋Š” IAPS ์ด๋ฏธ์ง€ ์„ธํŠธ์˜ Amusement์นดํ…Œ๊ณ ๋ฆฌ์—์„œ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ํŠน์ง•
    • classifier๋Š” ๊ธฐ๋ณธ์ ์œผ๋กœ ์ดˆ์ƒํ™”๋ฅผ ๊ฐ์ง€ํ•  ๋•Œ ๊ฐ€์žฅ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•จ
      • ๊ทธ๋Ÿฌ๋‚˜ artistic sets์˜ ์–ผ๊ตด๊ณผ ์นดํ…Œ๊ณ ๋ฆฌ ์‚ฌ์ด์—๋Š” ์ด๋Ÿฌํ•œ ๊ฐ•ํ•œ ์—ฐ๊ฒฐ์ด ์—†์Œ
        • ๋Œ€์‹  ์ƒ‰์ƒ์ด ํ›จ์”ฌ ๋” ์ค‘์š”ํ•ด์ง
          • ์˜ˆ์ˆ  ์ด๋ก (Itten colors, Wang Wei-ning histograms)์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ํŠน์ง•์ด ์‹ค์ œ๋กœ ์˜ˆ์ˆ ์  ์‚ฌ์ง„ ์„ธํŠธ์— ๊ฐ€์žฅ ์ž์ฃผ ์„ ํƒ๋จ
          • Amusement ๋ฐ Excitement: Itten features๊ฐ€ ๋งŽ์ด ์„ ํƒ๋จ, color features developed by Wang Wei-ning et al. ๋˜ํ•œ ํšจ๊ณผ์ ์ž„ (feature selection algorithms์— ์˜ํ•ด ์„ ํƒ๋˜์—ˆ์Œ)
          • Awe ๋ฐ Disgust: color features developed by Wang Wei-ning et al.์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€์Œ

โญ ์ •๋ฆฌํ•˜์ž๋ฉด

  • Anger ๋ฐ Contentment
    • IAPS + Yanulevskaya et al. features๊ฐ€ ๊ฐ€์žฅ ์ ํ•ฉ
  • Fear ๋ฐ Disgust
    • IAPS์˜ ๊ฒฝ์šฐ ์ฝ˜ํ…์ธ  (snakes, insects, injuries) ์˜ํ–ฅ์ด ํผ
  • Awe ๋ฐ Disgust
    • artistic sets + Wang Wei-ning histograms
  • Amusement ๋ฐ Excitement
    • IAPS์˜ ๊ฒฝ์šฐ ์ฝ˜ํ…์ธ  (ํ–‰๋ณตํ•œ ์‚ฌ๋žŒ์˜ ์ดˆ์ƒํ™”, occurrence์™€ ํฌ๊ธฐ) ์˜ํ–ฅ์ด ํผ
    • artistic sets + Itten colors/Wang Wei-ning histograms

 

CONCLUSIONS

  • ๋” ๋‚˜์€ ํŠน์ง•๊ณผ ๋” ๋‚˜์€ ๋ถ„๋ฅ˜๋ฅผ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์žˆ์–ด ๊ฐœ์„ ์˜ ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ์Œ
    • ์–ผ๊ตด์ด๋‚˜ ํŠน์ • ๊ณตํ†ต ๊ธฐํ˜ธ(e.g. ํ•˜ํŠธ)์˜ ๊ฐ์ • ํ‘œํ˜„ ์ธ์‹๊ณผ ๊ฐ™์€ semantic-based features๊ฐ€ ์œ ๋ฆฌํ•  ์ˆ˜ ์žˆ์Œ
    • ๋ถ„๋ฅ˜๋ฅผ ์œ„ํ•ด ๊ฐ ์ด๋ฏธ์ง€๋ฅผ ํ•˜๋‚˜์˜ ๊ฐ์ • ๋ฒ”์ฃผ์— ๊ฐ•์ œํ•˜๋Š” ๋Œ€์‹  ๋ฒ”์ฃผ์— ๋Œ€ํ•œ ๋ถ„ํฌ๋ฅผ ๋ณด์—ฌ์ฃผ๋Š” “emotional histogram”์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Œ
    • ๋” ๋งŽ์€ ์ˆ˜์˜ ์‚ฌ๋žŒ๋“ค๋กœ๋ถ€ํ„ฐ ๋” ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์‹ค์ธก ์ž๋ฃŒ๊ฐ€ ํ•„์š”
    • consensus-based approach(ํ•ฉ์˜ ๊ธฐ๋ฐ˜ ์ ‘๊ทผ๋ฒ•) ๋Œ€์‹  ๊ฐœ๋ณ„ ์‚ฌ๋žŒ๋“ค์˜ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ๋” ์ž ์žฌ์ ์ธ ๋ฐœ์ „์ž„