Food Image Classification

mxnet, Keras, resnet, resnext, ensemble learning

Transcript

1. ਓ޻஌ೳٕज़ઓུձٞओ࠵ AIνϟϨϯδίϯςετʢୈ1ճʣ (2)ྉཧ෼ྨ෦໳ɹϨϙʔτ ʮਂ૚ֶशΛ༻͍ͨը૾෼ྨʹ͓͚Δ Best Practiceʯ peroon 1
2. ɾσʔλͷ؍࡯ ɾσʔλͷར༻ํ๏ ɾԾઆ ɾϞσϦϯάͷ޻෉఺ ɾ༧ଌ݁Ռ͔ΒಘΒΕΔࣔࠦ ໨࣍ 2
3. bread_sandwich, bread_sliced, bread_sweets, bread_table, noodle_somen, noodle_udon, pasta_cream, pasta_gratin, pasta_japanese, pasta_oil, pasta_tomato, rice_boiled, rice_bowl, rice_curry, rice_fried, rice_risotto, rice_sushi, soup_miso, soup_potage, sweets_cheese, sweets_cookie, sweets_muffin, sweets_pie, sweets_pound, sweets_pudding σʔλͷ؍࡯ʢϥϕϧҰཡʣ 3 ࣅ͍ͯΔάϧʔϓ͸άϧʔϓԽͯ͠෼ྨ ͨ͠ํ͕͍͍ͷ͔ͳʁetc
4. σʔλͷ؍࡯ʢը૾ʣ 4 ثྖҬΛݕग़ͯ͠ɺͦͷத͚ͩΛֶशɾ෼ྨͨ͠ํ͕͍͍ͷ͔ͳʁ ը૾Ճ޻ʹΑΔςΩετ͸จࣈݕग़Ͱআڈͨ͠ํ͕͍͍ͷ͔ͳʁetc
5. ࠓճͷσʔλʹಛԽͨ͠ϧʔϧ ࡞Γ͸͍Ζ͍Ζߟ͑ΒΕΔ͕ɺ ͦΕΑΓ΋ɺCNNΛ༻͍ͨը૾ ෼ྨͷBest PracticeΛద༻ͯ͠ੑ ೳΛ࠷େʹҾ͖ग़͢͜ͱ͕ॏཁ Ծઆ 5
6. ɾԣ௕ͷը૾ʢαΠζ w x hʣͷ৔߹ɺw x wͷਖ਼ํܗ ը૾Λɺࠨ୺ɺதԝɺӈ୺͔Β੾ΓऔΔ ʢྉཧ͕ࠨӈʹدͬͯө͍ͬͯΔ΋ͷ͕͋ΔͨΊʣ ɾԣํ޲ʹॖখͯ͠w x wʹ͢Δ ʢ৘ใྔ͕Ұ൪࢒ΔͨΊʣ ɾ্هʹΑΓɺ1ຕͷը૾͕4ຕʹͳΔ ɾॎ௕ͷ৔߹ɺ੾ΓऔΓ͸্୺ɺதԝɺԼ୺ ɾ܇࿅ը૾ɺςετը૾ɺڞʹߦ͏ σʔλͷར༻ํ๏ʢϥϕϧ͋Γը૾ʣ 6
7. ࢖༻͠ͳ͔ͬͨ ɾը૾Λ؍࡯ͨ͠ͱ͜Ζɺ25Ϋϥεͱ͸ແؔ܎ͷը ૾΋ؚ·Ε͍ͯΔͨΊ ɾϥϕϧ͋Γը૾ͰCNNΛֶशޙɺϥϕϧͳ͠ը૾ Λ෼ྨͯ͠ԾͷϥϕϧΛ෇͚ͯ࠶ֶश(Pseudo Labeling)ͨ͠ͱ͜Ζɺvalidation accuracy͸্͕Δ ͕ɺtest accuracy͸มԽ͕ͳ͔ͬͨͨΊ σʔλͷར༻ํ๏ʢϥϕϧͳ͠ը૾ʣ 7
8. ɾimagenetͰֶशࡁΈͷϞσϧͷfine-tuning ɾfine tuning࣌ͷfreezeͤ͞Δ૚ͷਂ͞ͷௐ੔ ʢϋΠύʔύϥϝʔλͷRandom SearchͱCross ValidationͰܾఆʣ ɾvgg16, resnetͳͲҟͳΔϞσϧͷΞϯαϯϒϧֶश ɾը૾ͷcropping ɾ్தͷepoch͔Βlearning rateΛ0.1ഒʹ͢Δ ɾFC૚ʹDropoutΛೖΕΔ ɾ֤Ϟσϧ͸܇࿅σʔλͷ4/5ͷΈͰֶश͠ɺͦΕΛ5ͭ࡞Δ ɾը૾ͷલॲཧʢϞσϧ͝ͱʹܾ·͍ͬͯΔRGBฏۉΛҾ͘ʣ ɾ֤Ϟσϧ͕ϥϕϧΛ֬཰Ͱ༧ଌͨ͠΋ͷΛ߹ܭͯ͠࠷ऴ༧ଌ ɾ༧ଌ࣌ʹ୯ମͰ༏लͳϞσϧ(resnext)ͷॏΈΛ্͛Δ ɾ࠷ऴతʹ͸ɺresnext (101૚)Λ5Ϟσϧ࡞ͬͯΞϯαϯϒϧ ϞσϦϯάͷ޻෉఺ʢ্ख͍ͬͨ͘΋ͷʣ 8
9. ɾBatch Normalization ɾڧ͍Data Augmentation ɾֶश͕ઙ͍(epoch͕গͳ͍)CNNΛΞϯαϯϒϧʹؚΊΔ ɾNearest NeighbourΛΞϯαϯϒϧʹؚΊΔ ɾFC૚ʹSVM ɾϥϕϧͳ͠ը૾΁ͷPseudo Labeling ɾྉཧͷثݕग़ʢϋϑม׵ʹΑΔପԁݕग़ʣ ϞσϦϯάͷ޻෉఺ʢޮՌ͕ͳ͔ͬͨ΋ͷʣ 9
10. ɾثೝࣝͳͲࢥ͍͖ͭͷख๏͸ਫ਼౓޲্ʹد༩͠ͳ͔ͬͨҰ ํɺը૾ೝࣝͷ࿦จͰ࢖ΘΕ͍ͯΔख๏Λਖ਼͘͠࢖͏͜ͱͰ ண࣮ʹਫ਼౓͕޲্͠ɺϥϯΩϯάͰͷ࠷ߴਫ਼౓ʹ͍ۙ0.82Ҏ ্ͷείΞ͕ୡ੒Ͱ͖ͨɻԾઆͷ௨ΓɺBest Practice͕ޮ͘ͱ ͍͏͜ͱ͕ࣔࠦ͞ΕΔɻ ɾϏδωεͳͲͷ࣮Ԡ༻ʹ͓͍ͯ΋ɺ·ͣCNNͷੑೳΛे෼ ʹҾ͖ग़͠ɺͦͷޙʹ෼໺͝ͱͷϧʔϧΛՃ͍͑ͯ͘ͱ͍͏ ॱংͰਐΊΔ͜ͱ͕ɺண࣮ʹ໨ඪͱ͢Δਫ਼౓ʹۙͮͨ͘Ίͷ खஈͱͯ͠ਖ਼͍͜͠ͱ͕ࣔࠦ͞ΕΔɻ ༧ଌ݁Ռ͔ΒಘΒΕΔࣔࠦ 10