ΫοΫύουʹΔ Deep Learning Λ༻ྉཧըผͷऔΓΈ Approaches to Food/Non-food image classification using Deep Learning on cookpad ٠ ངฏ *1 Yohei Kikuta છ୩ ༔Ұ *1 Yuichiro Someya ϨγΣοΫ ϦϏπΩ *1 Leszek Rybicki *1 ΫοΫύουגձ Cookpad Inc. In this paper we report our approach to image classification, in particular to the food/non-food image classification problem, as used by our ྉཧΖ (Cooking Log) product of Cookpad Inc. We augment our existing services with a computationally expensive image analysis architecture implementing this solution. One challenge is that the non-food class is very vast and varied and can only be defined in context. We find that having the non-food class consist of multiple subclasses effectively improves both precision and recall by capturing different types of features in the images of the non-food class. 1. Ίʹ ຊߘͰ, ΫοΫύου *1 ʹΔʮྉཧΖʯ(ਤ 1) ͱ ϓϩμΫτΛࡐͱ, Deep Learning Λ༻ྉཧը ผʹΔऔΓΈΛใࠂΔ. ಛʹ,Deep Learning ΛΈࠐΜϓϩμΫτͷΞʔΩςΫνϟͷఏҊͱ, ྉཧɾඇ ྉཧผͷਫ਼ΛతͱࢪΞϓϩʔνͱͷ ՌͷߟΛߦ. ਤ 1: ྉཧΖͷΩϟϓνϟը໘. ࠨଆ Android Ͱӈଆ iOS ͷը໘ͰΓ, ܞଳதͷྉཧըΛಈతʹநग़ දΔ. ըΛλοϓΔͱͰϑΟʔυόοΫΛૹΔ ͱՄʹͳΔ. ྉཧΖϢʔβ 5 ສਓ, ྉཧͱผΕը 130 ສຕ (2017 3 7 ) ͷ ϓϩμΫτͰΔ. ਐలͷஶػցशͷͷதͰಛʹ Deep Learning ͷڻ୰ͷΔ, Deep Learning ΛΈࠐΜ ϓϩμΫτΛҰఆͷنͰఆৗతʹӡ༻Δ ଟͳͱࢥΘΕΔ. ͷཧ༝ͱҎԼͷΑͳͷ ߟΒΕΔ. ࿈བྷઌ: ٠ངฏ, ΫοΫύουגձ ڀݚ։෦, yohei- [email protected]*1 https://cookpad.com • ͳज़σʔλଗͳ ߴज़Λ MNIST ͷΑͳࡍͷαʔ Ϗεͱͷ࿈ബσʔλͰ׆ΔݶఆతͰ Δ. ·ɺσʔλ๛ͰΕΛ׆༻Δज़ͳ ΕػցशͷҖΛڗडΔͱͰͳ. • طଘαʔϏεʹΈࠐΉࡍͷίετϦεΫେ Deep Learning Λ༻ੳߴෛՙͰੳڥಛघ ͳͷΛཁٻΔΊ, ଞαʔϏεʹѱӨڹΛ༩ʹ༗ ʹѻʹੳҎ֎ͷݟͱज़ඞཁͰΔ. • අ༻ରՌࠐݟΊͳ ଟͷ߹,Deep Learning ΛऔΓೖΕʹߴՌΛΊͳ. தظతͳ࠲ࢹͱΕΛߦͰΔ ৫ମඞཁͰΔ. ΫοΫύουΕΒͷ՝ΛղαʔϏεΛఏڙͰΔ ك༗ͳۀاͷҰͰΓ, ࡏݱਐܗߦͰʑͳαʔϏεͷఏ ڙͱ৽ϓϩμΫτͷ։ΛਪਐΔ. ҎͷষͰ, զʑͷઃఆͱͷղͷΊʹࢪ۩ମతͳऔΓΈͷज़తଆ໘Λઆ, ͷՌͱޙࠓͷల Λड़Δ. 2. ઃఆ ΫοΫύουར༻ 6,300 ສਓͰϨγϐ 260 ສʢ2016 12 ʣͱຊ࠷େͷϨγϐ αʔϏεͰΔ. ଟͷϢʔβ༻ΔڊେͳαʔϏε ͰΔ, ڊେͰΔΏʹݶఆతͳར༻ʹ·ΔϢʔβ ଘࡏΔ. ϨγϐΛࡧݕྉཧΛ࡞ΓਅΔ, ཧతɾ৺ཧతোนʹΑΓ, ΕΛߘΔͱαʔϏεʹ ͱ༗༻ͳΞΫγϣϯ·Ͱ౸Βͳ߹গͳͳ. ΕαʔϏεʹͱॏཁͳ՝ͷҰͰΔ. ͷΑͳ՝ΛվળΔΊͷࢼΈͷҰͱ, ը ੳज़ʹجϓϩμΫτʹΔ. ۩ମతʹ, Ϣʔβͷ ܞଳதͷըΒྉཧըͷΈΛநग़, ྉཧͷཤ؆ ୯ʹӾཡͰΔͱʹՃ, ͷըΛىͱαʔϏεʹ ϑΟʔυόοΫૹΔͱͰΔʮྉཧΖʯͱϓϩ μΫτʹΔ. ຊߘͰ, ͷΑͳϓϩμΫτΛݱΔΊͷΞʔΩςΫνϟͷߏஙͱ, ྉཧըநग़෦ʹΔ 1 The 31st Annual Conference of the Japanese Society for Artificial Intelligence, 2017 1A1-OS-05a-1
4
Embed
7KH 1A1-OS-05a-1 « ¿ « Í ¿ Å t S Z DeepLearning ; M h g hþ ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
クックパッドにおけるDeep Learningを用いた料理画像判別の取り組み
Approaches to Food/Non-food image classification using Deep Learning on cookpad
菊田 遥平 ∗1
Yohei Kikuta
染谷 悠一郎 ∗1
Yuichiro Someya
レシェック リビツキ ∗1
Leszek Rybicki
∗1クックパッド株式会社Cookpad Inc.
In this paper we report our approach to image classification, in particular to the food/non-food image classificationproblem, as used by our 料理きろく (Cooking Log) product of Cookpad Inc. We augment our existing serviceswith a computationally expensive image analysis architecture implementing this solution. One challenge is that thenon-food class is very vast and varied and can only be defined in context. We find that having the non-food classconsist of multiple subclasses effectively improves both precision and recall by capturing different types of featuresin the images of the non-food class.
1. はじめに本稿では,クックパッド ∗1 における「料理きろく」(図 1)と
いうプロダクトを題材として, Deep Learningを用いた料理画像判別問題に関する取り組みを報告する. 特に,Deep Learning
実際の判別処理 (図の Bの部分)が非同期に行われることが肝要となる. 画像の判別結果や状態の管理等,判別処理以外の軽量な処理は API サーバーを介して行い, 判別処理そのものはメッセージキューを介することで非同期性を確保した上で行い, 判別結果を API サーバーを介してアプリケーションから利用可能な状態にする,という手法を採用した.