ルールベース画像処理のススメ

by fkubota

Slide 1

Slide 1 text

ϧʔϧϕʔεը૾ॲཧͷεεϝ fkubota DeepLearning ωΠςΟϒੈ୅΁޲͚ͯ

Slide 2

Slide 2 text

ࣗݾ঺հ 02 fkubota (Twitter, Kaggle) - όϯυϧΧʔυͷձࣾ(ΧϯϜ)ͰػցֶशΤϯδχΞ - Kaggle Expert - ԭೄग़਎(౦ژʹग़͖ͯͯ3೥൒) - ෺ཧֶՊग़਎(ڧ૬ؔిࢠܥɺ4.2KͷӷମϔϦ΢Ϝʹ࣓ੑମಥͬࠐΜͰͨ) - ϓϩάϥϛϯάྺ͸2೥൒͙Β͍ - झຯ - ૣى͖(4࣌൒ىচ) - Kaggle - ίʔώʔɺϏʔϧɺ΢ΟεΩʔ - ಡॻɺ෺ཧɺ఩ֶ

Slide 3

Slide 3 text

ͬͦ͘͞Ͱ͕͢ ͋Δݐ෺ͷϑϩΞը૾ ͜Εʹ ͜Μͳײ͡Ͱ੺͍఺Λଧ͍ͪͨ - ੨͍ྖҬʹೖͬͪΌͩΊ - ௨࿏͚ͩ Έͳ͞ΜͳΒͲ͏͠·͔͢ʁ 03

Slide 4

Slide 4 text

ൃදͷϞνϕʔγϣϯ ࣮͸͜ͷը૾ ͸Indoorίϯϖ ͷը૾Ͱ ੺͍఺ΛଧͭͨΊʹ؆୯ͳը૾ॲཧΛͯ͠ଧͪ·ͨ͠ νʔϜϝΠτʹղઆ͢Δͱ൓ԠΑ͔ͬͨͷͰ ϧʔϧϕʔεը૾ॲཧʹ͍ͭͯൃද͠Α͏ͱࢥ͍·ͨ͠ KaggleΛ͖͔͚ͬʹը૾ॲཧɾը૾ೝࣝΛ͸͡Ίͨਓ͸͍͖ͳΓ DeepLearning͔Βೖͬͨͱࢥ͍·͕͢ྺ࢙తʹ͸ϧʔϧϕʔε͸ͷ΄͏ ͕௕͘ɺະͩ׆༂͍ͯ͠Δٕज़Ͱ͢ɻ ֶΜͰଛ͸ͳ͍͸ͣʂʂ 04

Slide 5

Slide 5 text

ࠓ೔࿩͢͜ͱ σΟʔϓϥʔχϯάωΠςΟϒੈ୅ʹ޲͚ͯʂ 05 ը૾ͬͯͳʹʁ ը૾ॲཧೖ໳ Indoorͷ৔߹ ͦ΋ͦ΋ը૾ͬͯͳʹ͔Λ੔ ཧ͠·͢ɻ ϧʔϧϕʔεը૾ॲཧʹͪΐͬ ͱ͚ͩೖ໳ͯ͠Έ·͢ɻ opencvΛ࢖͍·͢ɻ https://github.com/fkubota/ bunseki_compe_LT_02/blob/main/ notebook/nb01_opencv.ipynb IndoorͰ΍ͬͨ͜ͱΛ΄Μͷগ ͠঺հ͠·͢ɻ

Slide 6

Slide 6 text

ը૾ͬͯԿʁ

Slide 7

Slide 7 text

ը૾Λ؆୯ʹ - pixel ͱ͍͏࠷খ୯ҐΛ΋ͭ - ࠲ඪx, y Λ΋ͭ - ً౓஋Λ΋ͭ - ௨ৗ͸8bit - 0~255ͷ஋ΛऔΔ(256ஈ֊) - 0͕ࠇ - 255͕ന y x നͬΆ͍ͷͰ 255ʹ͍ۙ஋ ࠇͳͷͰ0ʹ͍ۙ஋ փ৭ͳͷͰਅΜதͷ 150͙Β͍ʁ ΧϥʔͩͱRGBͷ3͕ͭͦΕͧΕ8bitͷ৘ใΛ࣋ͭ ྫ: - (133, 30, 88) ←10ਐ਺ - #56a53f ←16ਐ਺ 07

Slide 8

Slide 8 text

ը૾ॲཧೖ໳

Slide 9

Slide 9 text

ϧʔϧϕʔεͷ͍͍ͱ͜Ζ ڭࢣσʔλ͍Βͳ͍ ָ͍͠ʂ - ΍ͬͨΒΘ͔Δʂʂʂʂ - ͜Εҙ֎ͱݟམͱ͞Ε͕ͪ 09

Slide 10

Slide 10 text

໰1 from fkubota shibutsu dataset Έͳ͞ΜͳΒͲ͏͠·͔͢ʁ (ڭࢣσʔλͳ͠Ͱ͢Αʁ) Ωϟοϓͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ ༡ͼײ֮Ͱʂʂ 10 https://github.com/fkubota/ bunseki_compe_LT_02/blob/main/ notebook/nb01_opencv.ipynb શ෦githubʹ͋ΔΑ

Slide 11

Slide 11 text

ղ๏ͷ1ྫΛ঺հ ·ͣ͸ఆ൪ͷ2஋ԽΛ͢ΔͨΊʹ άϨʔը૾ʹม׵͠·͢ʂ ֤pixelͷً౓஋(r, g, b) ͷ3࣍ݩ͕ 1࣍ݩʹͳΓ͍Ζ͍Ζѻ͍΍͘͢ͳΓ·͢ɻ 11

Slide 12

Slide 12 text

2஋Խ എܠͱ෺ମΛ෼͚·͠ΐ͏ ً౓஋ͷώετάϥϜΛݟͯΈΔͱ - നͬΆ͍৭(255ʹ͍ۙ৭)͕ώετάϥ ϜͷϚδϣϦςΟʹͳͬͯΔ - ͜Ε͸എܠ͕ը૾ͷେ෦෼Λ઎ΊΔ͜ ͱʹىҼ͢Δ - threshold = 70 ͷલޙͰ2஋Խ͢Δʂ 1 0 Ωϟοϓ എܠ 12

Slide 13

Slide 13 text

2஋Խ݁Ռ ϐϯΫͷ෦෼͕1 ͦΕҙ֎͕0 ͷ ΦϒδΣΫτ ͕Ͱ͖ͨ ͜ΕΛ “ྖҬ” ͱݴ͍·͢ 13

Slide 14

Slide 14 text

͋ͱ͸៉ྷʹ͢Δ ࠓճ͸݀ΛຒΊ͚ͨͩ 14

Slide 15

Slide 15 text

ॏ৺Λܭࢉ ྖҬͷྠֲΛऔͬͯॏ৺Λܭࢉ Ͱ͖ͨʂʂ 15

Slide 16

Slide 16 text

໰2 from fkubota shibutsu dataset Ωϟοϓͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ ָ͘͠ͳ͖ͬͯ·ͨ͠ʁ 16

Slide 17

Slide 17 text

Կ͕೉͍͠ʁ Ωϟοϓͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ ಉ͡໰୊ͳΜ͚ͩͲෳ਺෺ମ͕͋ΔͷͰ ೉қ౓͕௓Ͷ্͕Γ·͢Ͷɻ 17

Slide 18

Slide 18 text

ఆ൪ͷॲཧ άϨʔը૾ 2஋Խ 18

Slide 19

Slide 19 text

νϣοτ͖Ε͍ʹ ϞϧϑΥϩδʔม׵ - ࣌ؒͷ౎߹্ৄࡉ͸ল͖·͢ - ϊΠζআڈɺ݀ຒΊ͕Ͱ͖·͢ - ݪཧ͸؆୯ͳͷͰͥͻௐ΂ͯΈͯ ͍ͩ͘͞ ݀ຒΊ ΊͬͪΌ͖Ε͍ʹͳͬͨʂʂ 19

Slide 20

Slide 20 text

ϥϕϦϯά ࿈ଓͨ͠ΦϒδΣΫτຖʹ෼ׂ - ྡΓ߹ͬͨ1ΛಉҰͷϥϕϧͱͯ͠෼ׂ͢Δ - ϐϯΫͷྖҬ͕ෳ਺ͷྖҬʹ෼ׂ͞Εͨ - ෼ׂޙ໘ੵ͕খ͍͞΋ͷΛϊΠζͱͯ͠আ͍ͨ ໘ੵͷখ͍͞ྖҬ(ϊΠζ) 20

Slide 21

Slide 21 text

͓ΘΓ͡Όͳ͍Αʂ ͜ͷத͔ΒΩϟοϓΛબͼ ग़͞ͳ͍ͱ͍͚·ͤΜ ಛ௃ྔΛ࢖͍·͢ ͜ͷ5ͭͷ෺ମΛൺֱͯ͠Ωϟοϓʹ͸Ͳ͏ ͍ͬͨಛ௃͕͋ΔͰ͠ΐ͏͔ʁ - ؙΈ͕͋Δɹ - ໘ੵ͕খ͍͞ ͕ࢥ͍౰ͨΓ·͢ɻ ໘ੵ͕খ͍͞Λ࢖ͬͯΈ·͢ɻ ͭ·Γ ಛ௃ྔʹ໘ੵΛ࢖͏ͱ͍͏͜ͱʹͳΓ·͢ɻ 21

Slide 22

Slide 22 text

ྖҬͷ໘ੵΛܭࢉ ໘ੵ࠷খΛબͿ ಛ௃ྔ = ໘ੵ Ͱ͖ͨʂʂ 22

Slide 23

Slide 23 text

໰3 Ͱ͖ͦʁ from fkubota shibutsu dataset ϖϯͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ 23

Slide 24

Slide 24 text

ϥϕϦϯά·Ͱ͸Ұॹ ϖϯͷಛ௃Λߟ͑Δ fkubota͞Μ͸ɺʮࡉ௕͍ʯҙ֎ ࢥ͍͖ͭ·ͤΜͰͨ͠ɻ ࡉ௕͍͸͜͏දݱ͢Δ͜ͱʹ͠·͢ɻ 1. ྖҬΛ࠷খ֎઀ۣܗͰғ͏ 2. ʮ௕ลͷ௕͞ʯΛʮ୹ลͷ௕͞ʯͰׂΔ 3. ͜ΕΛʮࡉ௕͞౓ʯͱ͢Δ 24

Slide 25

Slide 25 text

ࡉ௕͍౓Λܭࢉ ಛ௃ྔ = ࡉ௕͍౓ ࠷େ஋ΛબͿ Ͱ͖ͨʂʂ 25

Slide 26

Slide 26 text

໰4 ࠷ޙͩΑʂ from fkubota shibutsu dataset MINTIAͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ 26

Slide 27

Slide 27 text

ಛ௃͸ʁ MINTIAͷಛ௃Λߟ͑Δ ࢛͍֯ͱ͍͏ಛ௃͕͋Γ·͕͢ɺ ଞͷ෺ମͱࠩผԽ͕೉͍͠Ͱ͢ɻ ܗͷ؍఺Ͱ͸೉ͦ͠͏ͳͷͰ ʮ৭ʯͰ߈ΊͯΈ·͠ΐ͏ɻ ʮ੨͞ʯΛදݱ͠·͢ɻ RGBΛHSVʹม׵͠ H(hue, ৭૬)ͷը૾Λ࢖͍·͢ɻ 27

Slide 28

Slide 28 text

৭૬ʁ ͜ͷล͕੨ ৭૬͸৭ͷ༷૬Λ 0~180·ͰͰมԽͤͨ͞΋ͷɻ ੨͸100~140͋ͨΓɻ 28

Slide 29

Slide 29 text

Hueը૾ hueը૾ ͜ͷลΓ 29

Slide 30

Slide 30 text

2஋Խͯ͠ϞϧϑΥϩδʔ 2஋Խ ϞϧϑΥϩδʔ 30

Slide 31

Slide 31 text

ॏ৺Λܭࢉͯ͠ऴΘΓʂ Ͱ͖ͨʂʂ 31

Slide 32

Slide 32 text

Indoorͷ৔߹

Slide 33

Slide 33 text

࠷ॳͷ࣭໰ ͋Δݐ෺ͷϑϩΞը૾ ͜Εʹ ͜Μͳײ͡Ͱ੺͍఺Λଧ͍ͪͨ - ੨͍ྖҬʹೖͬͪΌͩΊ - ௨࿏͚ͩ Έͳ͞ΜͳΒͲ͏͠·͔͢ʁ 33

Slide 34

Slide 34 text

·ͣ͸௨࿏Λऔಘ ੨͍෦෼Λ2஋Խ ຒΊΔ ന͍෦෼Λ 2஋Խ ࿦ཧੵ 34