画像認識：仕組みと応用例

コンピュータサイエンス・アドベンチャー　～理論計算機科学はこんなに面白い！～
2022年11月20日

Transcript

1. ը૾ೝࣝ ࢓૊ΈͱԠ༻ྫ ۄ໦ప ʢ໊ݹ԰޻ۀେֶ޻ֶ෦৘ใ޻ֶՊʣ ίϯϐϡʔλαΠΤϯεɾΞυϕϯνϟʔ ʙཧ࿦ܭࢉػՊֶ͸͜Μͳʹ໘ന͍ʂʙ 2022೥11݄20೔ https://bit.ly/20221120tamaki
2. ը૾ೝࣝͷԠ༻ྫ
3. εϚϗɼө૾ฤू Pixabay License https://pixabay.com/ja/photos/iphone-εϚʔτ- ϑΥϯ-410311/ Pixabay License https://pixabay.com/ja/photos/ਓؒ- photoshop-4896511/
4. ؂ࢹΧϝϥ Pixabay License https://pixabay.com/ja/photos/Χϝϥ-؂ࢹ-ϏσΦ؂ࢹ-2323373/ Pixabay License https://pixabay.com/ja/photos/cctv-؂ࢹ-Χϝϥ-2805301/
5. https://dempa- digital.com/article/178588 2021/3/31
6. ޻৔ Siyuwj - Own work CC BY-SA 3.0 https://commons.wikimedia.org/wiki/File:Geely_assembly_line_in_Beilun,_Ningbo.JPG CC0 https://www.pikrepo.com/fciko/assembly-line-machine-at-maker-s-mark- distillery
7. ҩ༻ը૾ daveynin from United States - New UPMC East CC BY 2.0 https://commons.wikimedia.org/wiki/File:UPMCEast_CTscan.jpg Pixabay License https://pixabay.com/ja/illustrations/ίϯϐϡʔλʔஅ૚ࡱӨ-ct-62942/ By Gilo1969 - Own work, CC BY 2.0 https://commons.wikimedia.org/wiki/Fi le:Endoscopy_room.jpg
8. ࣗಈӡసɼंࡌΧϝϥө૾ Eschenzweig - Own work CC BY-SA 4.0 https://commons.wikimedia.org/wiki/File:Autonomous-driving-Barcelona.jpg Steve Jurvetson derivative work: Mariordo CC BY 2.0 https://ja.wikipedia.org/wiki/ϑΝΠϧ:Google%27s_Lexus_RX_450h_Self-Driving_Car.jpg
9. Ӵ੕ը૾ɼߤۭࣸਅ NASA/Apollo 17 crew Public domain https://commons.wikimedia.org/wiki/File:The_Earth_seen_from_Apollo_17.jpg NASA Hubble CC BY 2.0 https://commons.wikimedia.org/wiki/File:Grappling_the_Hubble_Space_Telescope.jpg
10. https://prtimes.jp/main/html/rd/p/ 000000018.000026963.html 2020/8/4
11. ؾ৅Ӵ੕ը૾ She-Hulka - ESODIS Worldwide CC BY-SA 4.0 https://commons.wikimedia.org/wiki/File:Japan_on_a_satellite.jpg NASA Public domain https://commons.wikimedia.org/wiki/File:ParmaMelor_AMO_TMO_2009279_lrg.jpg
12. https://ledge.ai/u-ryukyu-ac-typhoon/ 2021/6/29
13. ը૾ೝࣝͷԠ༻ྫʹؔ͢ΔهࣄϦϯΫ https://raindrop.io/ttta maki/-11603202
14. ػցֶशͷجຊతͳ࢓૊Έ ڭࢣ͋Γֶश
15. ڭࢣ͋Γֶश ʜ ʜ ʜ ʜ ֶशσʔληοτ ڭࢣϥϕϧ             ςετσʔλ ࣝผ  ֶशαϯϓϧ ֶश ࣝผث
16. ڭࢣ͋Γֶशɿࣝผͱճؼ nೖग़ྗ • ೖྗxɿԿΒ͔ͷσʔλʢը૾ɼԻ੠ɼηϯαɼςΩετɼɽɽɽʣ • ग़ྗyɿ • ࣝผɿ཭ࢄΧςΰϦʢछྨɼϥϕϧɼΧςΰϦɼΫϥεʣ • ճؼɿ࿈ଓ஋ nग़ྗͷछྨ • ࣝผ • ΧςΰϦ਺ʢΫϥε਺ʣ͕༗ݶݸ • ཭ࢄత • ճؼ • ਺஋ɼ࿈ଓత
17. ը૾ೝࣝɼը૾ࣝผ nإೝࣝ • ೖྗɿը૾ • ग़ྗɿإɼͦΕҎ֎ nࢦ໲ೝূ • ೖྗɿࢦ໲ը૾ • ग़ྗɿొ࿥ऀA, B, C, …., ඇొ࿥ nΨϯೝࣝ • ೖྗɿҩྍσʔλ • ग़ྗɿΨϯɼඇΨϯ nδΣενϟೝࣝ • ೖྗɿηϯα৴߸ • ग़ྗɿొ࿥δΣενϟͷछྨ CC BY-SA 3.0 The Photographer Wilfredor By Gilo1969 - Own work, CC BY 2.0
18. ςΩετೝࣝɼԻ੠ೝࣝ n໎࿭ϝʔϧϑΟϧλ • ೖྗɿϝʔϧ • ग़ྗɿී௨ͷϝʔϧɼ໎࿭ϝʔϧ nυΩϡϝϯτ෼ྨ • ೖྗɿςΩετ • ग़ྗɿχϡʔεɼεϙʔπɼܳೳɼɽɽɽ nԻ੠ೝࣝ • ೖྗɿԻ੠৴߸ • ग़ྗɿԻૉɼςΩετ ͜Μʹͪ͸ʢLPOOJDIJXBʣ
19. ճؼ nגՁ༧ଌ • ೖྗɿաڈͷגՁɼגࣜσʔλ • ग़ྗɿকདྷͷגՁɼ্͕Δ͔Լ͕Δ͔ nՁ֨༧ଌ • ೖྗɿՁ֨ɼ༷ʑͳσʔλ • ग़ྗɿՁ֨ nإ೥ྸਪఆ • ೖྗɿը૾ • ग़ྗɿ೥ྸ ʁ ʁ CC BY 3.0 By Monaneko - http://www.stat.go.jp/data/getujidb/zuhyou/d09.xls, GFDL, https://commons.wikimedia.org/w/index.php?curid=2412825 By Tosaka - Own work, CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=3111258   
20. ը૾ೝࣝͷجຊతͳ࢓૊Έ ৞ΈࠐΈχϡʔϥϧωοτϫʔΫ
21. $//ɿ৞ΈࠐΈχϡʔϥϧωοτϫʔΫ Aphex34 Own work CC BY-SA 4.0 https://commons.wikimedia.org/wiki/File:Typical_cnn.png ը૾ ϥϕϧ
22. $//ͷΦϯϥΠϯσϞ https://transcranial. github.io/keras- js/#/mnist-cnn
23. $//ɿೖྗYɼग़ྗZ Aphex34 Own work CC BY-SA 4.0 https://commons.wikimedia.org/wiki/File:Typical_cnn.png ը૾ ϥϕϧ ೖྗ ग़ྗ ೣ cats and dogs dataset Ξϊςʔλʔ ϥϕϧ෇͚ ࡞ۀ
24. ༷ʑͳը૾ೝࣝɾը૾ॲཧ
25. ը૾ೝࣝɿೖྗը૾ɼग़ྗϥϕϧ ೣ cats and dogs dataset ݘ
26. CIFAR-10 / CIFAR-100 https://www.cs.toronto.edu/~kriz/cifar.html MNIST http://yann.lecun.com/exdb/mnist/ https://commons.wikimedia.org/wiki/File:MnistExamples.png
27. ৄࡉը૾ೝࣝɿೖྗը૾ɼग़ྗϥϕϧ Abyssinian ΞϏγχΞϯ Oxford-IIIT-Pet dataset american bulldog ΞϝϦΧϯϒϧυοά basset hound όηοτϋ΢ϯυ Maine Coon ϝΠϯΫʔϯ
28. Caltech-UCSD Birds-200-2011 http://www.vision.caltech.edu/visipedia/CUB-200-2011.html Stanford Cars dataset https://ai.stanford.edu/~jkrause/cars/car_dataset.html The Oxford-IIIT Pet Dataset https://www.robots.ox.ac.uk/~vgg/data/pets/
29. ྖҬ෼ׂɿೖྗը૾ɼग़ྗը૾ ը૾ ϥϕϧը૾ https://cocodataset.or g/workshop/coco- lvis-eccv-2020.html Joint COCO and LVIS Recognition Challenge Workshop at ECCV 2020
30. Cameron Laedtke, Semantic Segmentation Demo: SegFormer B2 https://www.youtube.com/watch?v=UJLLWhdiCG4
31. ෺ମݕग़ɿೖྗը૾ɼग़ྗ෺ମ৘ใʢۣܗʣ (MTheiler) - Own work CC BY-SA 4.0 https://commons.wikimedia.org/wiki/File:Detected-with-YOLO--Schreibtisch-mit- Objekten.jpg#/media/File:Detected-with-YOLO--Schreibtisch-mit-Objekten.jpg ը૾ • ෺ମ1ͷΧςΰϦ • ෺ମ1ͷۣܗ࠲ඪʢx1, y1, x2, y2ʣ • … • … • … • ෺ମ10ͷΧςΰϦ • ෺ମ10ͷۣܗ࠲ඪʢx1, y1, x2, y2ʣ
32. https://modeldepot.github.io/tfj s-yolo-tiny-demo/
33. ࢟੎ਪఆɿೖྗը૾ɼग़ྗ࢟੎৘ใʢࠎ֨ʣ ը૾ ࢟੎৘ใ https://github.com/CMU-Perceptual-Computing-Lab/openpose
34. https://teachablemachine.withgoogle.com/train
35. ܈ऺΧ΢ϯτɿೖྗը૾ɼग़ྗ਺ ը૾ ਓ਺ʢີ౓Ϛοϓʣ ShanghaiTech Dataset https://github.com/desenzhou/ShanghaiTechDataset Single-Image Crowd Counting via Multi-Column Convolutional Neural Network, CVPR2016
36. આ໌จੜ੒ɿೖྗը૾ɼग़ྗςΩετ a big group of people riding horses through the woods. a group of horse riders following a trail. a group of people riding horse through a forest on a dirt path. a group of horseback riders on a trail in the woods. a group of people riding horses down a trail. http://cocodataset.org/#explore?id=162252 MS-COCO dataset
37. http://captions.stair.center/demo/
38. 72"ɿೖྗը૾ςΩετɼग़ྗςΩετ౳ What is the mustache made of ? banana How many people can fit in the 2 buses? 40, 80, 100, 100, 100, 100, 200, many, many, lot VQAv2 dataset
39. https://vilbert.cloudcv.org
40. ը૾ม׵ͷ༷ʑͳྫ
41. ը૾ม׵ɿೖྗઢըɼग़ྗ࣮ը૾
42. https://affinelayer.com/pixsrv/
43. ը૾ͷΧϥʔԽɿೖྗനࠇը૾ɼग़ྗΧϥʔը૾ ௨ৗͷը૾ॲཧ Oxford-IIIT-Pet dataset
44. http://iizuka.cs.tsukuba.ac.jp/projects/colorization/web/
45. ௒ղ૾ɿೖྗ௿ղ૾౓ը૾ɼग़ྗߴղ૾౓ը૾ ௨ৗͷը૾ॲཧ
46. Image Super-Resolution via Iterative Refinement https://arxiv.org/pdf/2104.07636.pdf Google͕ʮΨϏΨϏͷ௿ղ૾౓ը૾Λߴղ૾౓ը૾ʹม׵͢ΔAIϞσϧʯͷੑೳΛվળɺਓ͕ؒ൑ผͰ͖ͳ͍Ϩϕϧʹ https://gigazine.net/news/20210831-google-upscaling-low-resolution-images/
47. ΠϯϖΠϯςΟϯάɿೖྗ͖݀͋ը૾ɼग़ྗิ׬ը૾ ௨ৗͷը૾ॲཧ
48. https://apps.apple.com/jp/app/phot o-retouch-ը૾Ճ޻ࣸਅਓΛফ͢ /id1230394683
49. ϘέআڈɿೖྗϘέͨը૾ɼग़ྗγϟʔϓͳը૾ ௨ৗͷը૾ॲཧ
50. ϊΠζআڈɿೖྗϊΠζ͋Γը૾ɼग़ྗΩϨΠͳը૾ ௨ৗͷը૾ॲཧ
51. ը૾ೝ͕ࣝ࣋ͭ՝୊
52. ϥϕϧϊΠζɿਖ਼͘͠ͳ͍ϥϕϧ෇͚ Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks, NeurIPS2021 https://openreview.net/forum?id=XccDXrDNLek
53. https://labelerrors.com
54. ఢରత߈ܸɿը૾ೝࣝͷηΩϡϦςΟ໰୊ Explaining and Harnessing Adversarial Examples, ICLR2015 https://arxiv.org/abs/1412.6572 ύϯμ ख௕Ԑ <latexit sha1_base64="ZVNwg01MO2fFIzdwROknkyHvcMs=">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</latexit> xi + xi
55. https://kennysong.git hub.io/adversarial.js/
56. ·ͱΊ nը૾ೝࣝͷ࢓૊Έ • ೖྗxΛड͚औͬͯग़ྗyΛग़ؔ͢਺f • σʔληοτ(x, y)͕ඞཁ • ग़ྗyΛΞϊςʔγϣϯ͢Δ • ೖྗxΛΞϊςʔγϣϯ͢Δ • ग़ྗy͔ΒೖྗxΛੜ੒ͯ͠͠·͏ nը૾ೝࣝͷ՝୊ • ग़ྗyͷΞϊςʔγϣϯ͕৴པͰ͖Δ͔ • ೖྗxͷมಈʹڧ͍͔
57. ৯ࣄॲͷ͝Ҋ಺