データ分析LT会第二回で発表した際の資料です。 youtube: https://www.youtube.com/watch?v=jDZwX3jxhK4
conppass url: https://kaggle-friends.connpass.com/event/214854/
github repository: https://github.com/fkubota/bunseki_compe_LT_02
ϧʔϧϕʔεը૾ॲཧͷεεϝfkubotaDeepLearning ωΠςΟϒੈ͚ͯ
View Slide
ࣗݾհ 02fkubota (Twitter, Kaggle)- όϯυϧΧʔυͷձࣾ(ΧϯϜ)ͰػցֶशΤϯδχΞ- Kaggle Expert- ԭೄग़(౦ژʹग़͖ͯͯ3)- ཧֶՊग़(ڧ૬ؔిࢠܥɺ4.2KͷӷମϔϦϜʹ࣓ੑମಥͬࠐΜͰͨ)- ϓϩάϥϛϯάྺ2͙Β͍- झຯ- ૣى͖(4࣌ىচ)- Kaggle- ίʔώʔɺϏʔϧɺΟεΩʔ- ಡॻɺཧɺֶ
ͬͦ͘͞Ͱ͕͋͢ΔݐͷϑϩΞը૾͜Εʹ ͜Μͳײ͡Ͱ͍Λଧ͍ͪͨ- ੨͍ྖҬʹೖͬͪΌͩΊ- ௨࿏͚ͩΈͳ͞ΜͳΒͲ͏͠·͔͢ʁ03
ൃදͷϞνϕʔγϣϯ࣮͜ͷը૾ Indoorίϯϖ ͷը૾Ͱ͍ΛଧͭͨΊʹ؆୯ͳը૾ॲཧΛͯ͠ଧͪ·ͨ͠νʔϜϝΠτʹղઆ͢ΔͱԠΑ͔ͬͨͷͰϧʔϧϕʔεը૾ॲཧʹ͍ͭͯൃද͠Α͏ͱࢥ͍·ͨ͠KaggleΛ͖͔͚ͬʹը૾ॲཧɾը૾ೝࣝΛ͡Ίͨਓ͍͖ͳΓDeepLearning͔Βೖͬͨͱࢥ͍·͕͢ྺ࢙తʹϧʔϧϕʔεͷ΄͏͕͘ɺະͩ׆༂͍ͯ͠Δٕज़Ͱ͢ɻֶΜͰଛͳ͍ͣʂʂ04
ࠓ͢͜ͱ σΟʔϓϥʔχϯάωΠςΟϒੈʹ͚ͯʂ05ը૾ͬͯͳʹʁ ը૾ॲཧೖ Indoorͷ߹ͦͦը૾ͬͯͳʹ͔Λཧ͠·͢ɻϧʔϧϕʔεը૾ॲཧʹͪΐͬͱ͚ͩೖͯ͠Έ·͢ɻopencvΛ͍·͢ɻhttps://github.com/fkubota/bunseki_compe_LT_02/blob/main/notebook/nb01_opencv.ipynbIndoorͰͬͨ͜ͱΛ΄Μͷগ͠հ͠·͢ɻ
ը૾ͬͯԿʁ
ը૾Λ؆୯ʹ- pixel ͱ͍͏࠷খ୯ҐΛͭ- ࠲ඪx, y Λͭ- ًΛͭ- ௨ৗ8bit- 0~255ͷΛऔΔ(256ஈ֊)- 0͕ࠇ- 255͕നyxനͬΆ͍ͷͰ255ʹ͍ۙࠇͳͷͰ0ʹ͍ۙփ৭ͳͷͰਅΜதͷ150͙Β͍ʁΧϥʔͩͱRGBͷ3͕ͭͦΕͧΕ8bitͷใΛ࣋ͭྫ:- (133, 30, 88) ←10ਐ- #56a53f ←16ਐ07
ը૾ॲཧೖ
ϧʔϧϕʔεͷ͍͍ͱ͜Ζڭࢣσʔλ͍Βͳ͍ ָ͍͠ʂ- ͬͨΒΘ͔Δʂʂʂʂ- ͜Εҙ֎ͱݟམͱ͞Ε͕ͪ09
1from fkubota shibutsu datasetΈͳ͞ΜͳΒͲ͏͠·͔͢ʁ(ڭࢣσʔλͳ͠Ͱ͢Αʁ)Ωϟοϓͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞༡ͼײ֮Ͱʂʂ10https://github.com/fkubota/bunseki_compe_LT_02/blob/main/notebook/nb01_opencv.ipynbશ෦githubʹ͋ΔΑ
ղ๏ͷ1ྫΛհ·ͣఆ൪ͷ2ԽΛ͢ΔͨΊʹάϨʔը૾ʹม͠·͢ʂ֤pixelͷً(r, g, b) ͷ3࣍ݩ͕1࣍ݩʹͳΓ͍Ζ͍Ζѻ͍͘͢ͳΓ·͢ɻ11
2ԽഎܠͱମΛ͚·͠ΐ͏ًͷώετάϥϜΛݟͯΈΔͱ- നͬΆ͍৭(255ʹ͍ۙ৭)͕ώετάϥϜͷϚδϣϦςΟʹͳͬͯΔ- ͜Εഎܠ͕ը૾ͷେ෦ΛΊΔ͜ͱʹىҼ͢Δ- threshold = 70 ͷલޙͰ2Խ͢Δʂ1 0Ωϟοϓഎܠ12
2Խ݁ՌϐϯΫͷ෦͕1ͦΕҙ֎͕0ͷ ΦϒδΣΫτ ͕Ͱ͖ͨ͜ΕΛ “ྖҬ” ͱݴ͍·͢13
͋ͱ៉ྷʹ͢Δࠓճ݀ΛຒΊ͚ͨͩ14
ॏ৺ΛܭࢉྖҬͷྠֲΛऔͬͯॏ৺ΛܭࢉͰ͖ͨʂʂ15
2from fkubota shibutsu datasetΩϟοϓͷ࠲ඪΛऔಘָ͍ͯͩ͘͘͠͞͠ͳ͖ͬͯ·ͨ͠ʁ16
Կ͕͍͠ʁΩϟοϓͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞ಉ͡ͳΜ͚ͩͲෳମ͕͋ΔͷͰқ͕Ͷ্͕Γ·͢Ͷɻ17
ఆ൪ͷॲཧάϨʔը૾ 2Խ18
νϣοτ͖Ε͍ʹϞϧϑΥϩδʔม- ࣌ؒͷ߹্ৄࡉল͖·͢- ϊΠζআڈɺ݀ຒΊ͕Ͱ͖·͢- ݪཧ؆୯ͳͷͰͥͻௐͯΈ͍ͯͩ݀͘͞ຒΊΊͬͪΌ͖Ε͍ʹͳͬͨʂʂ19
ϥϕϦϯά࿈ଓͨ͠ΦϒδΣΫτຖʹׂ- ྡΓ߹ͬͨ1ΛಉҰͷϥϕϧͱׂͯ͢͠Δ- ϐϯΫͷྖҬ͕ෳͷྖҬʹׂ͞Εͨ- ׂޙ໘ੵ͕খ͍͞ͷΛϊΠζͱͯ͠আ͍ͨ໘ੵͷখ͍͞ྖҬ(ϊΠζ)20
͓ΘΓ͡Όͳ͍Αʂ͜ͷத͔ΒΩϟοϓΛબͼग़͞ͳ͍ͱ͍͚·ͤΜಛྔΛ͍·͢͜ͷ5ͭͷମΛൺֱͯ͠ΩϟοϓʹͲ͏͍ͬͨಛ͕͋ΔͰ͠ΐ͏͔ʁ- ؙΈ͕͋Δɹ- ໘ੵ͕খ͍͕͞ࢥ͍ͨΓ·͢ɻ໘ੵ͕খ͍͞ΛͬͯΈ·͢ɻͭ·Γಛྔʹ໘ੵΛ͏ͱ͍͏͜ͱʹͳΓ·͢ɻ21
ྖҬͷ໘ੵΛܭࢉ໘ੵ࠷খΛબͿಛྔ = ໘ੵͰ͖ͨʂʂ22
3 Ͱ͖ͦʁfrom fkubota shibutsu datasetϖϯͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞23
ϥϕϦϯά·ͰҰॹϖϯͷಛΛߟ͑Δfkubota͞Μɺʮࡉ͍ʯҙ֎ࢥ͍͖ͭ·ͤΜͰͨ͠ɻࡉ͍͜͏දݱ͢Δ͜ͱʹ͠·͢ɻ1. ྖҬΛ࠷খ֎ۣܗͰғ͏2. ʮลͷ͞ʯΛʮลͷ͞ʯͰׂΔ3. ͜ΕΛʮࡉ͞ʯͱ͢Δ24
ࡉ͍Λܭࢉಛྔ = ࡉ͍ ࠷େΛબͿͰ͖ͨʂʂ25
4 ࠷ޙͩΑʂfrom fkubota shibutsu datasetMINTIAͷ࠲ඪΛऔಘ͍ͯͩ͘͠͞26
ಛʁMINTIAͷಛΛߟ͑Δ࢛͍֯ͱ͍͏ಛ͕͋Γ·͕͢ɺଞͷମͱࠩผԽ͕͍͠Ͱ͢ɻܗͷ؍Ͱͦ͠͏ͳͷͰʮ৭ʯͰ߈ΊͯΈ·͠ΐ͏ɻʮ੨͞ʯΛදݱ͠·͢ɻRGBΛHSVʹม͠H(hue, ৭૬)ͷը૾Λ͍·͢ɻ27
৭૬ʁ͜ͷล͕੨৭૬৭ͷ༷૬Λ0~180·ͰͰมԽͤͨ͞ͷɻ੨100~140͋ͨΓɻ28
Hueը૾hueը૾͜ͷลΓ29
2Խͯ͠ϞϧϑΥϩδʔ2Խ ϞϧϑΥϩδʔ30
ॏ৺Λܭࢉͯ͠ऴΘΓʂͰ͖ͨʂʂ31
Indoorͷ߹
࠷ॳͷ࣭͋ΔݐͷϑϩΞը૾͜Εʹ ͜Μͳײ͡Ͱ͍Λଧ͍ͪͨ- ੨͍ྖҬʹೖͬͪΌͩΊ- ௨࿏͚ͩΈͳ͞ΜͳΒͲ͏͠·͔͢ʁ33
·ͣ௨࿏Λऔಘ੨͍෦Λ2ԽຒΊΔന͍෦Λ2Խཧੵ34
ʂ&͍ʑΛॻ͍ͨը૾Λ༻ҙͯ͠ཧੵͰ͖ͨʂʂ35
Thanks :)