System is processing data
Please download to view
...

ICIP2014 Presentation

by takayoshi-yamashita

on

Report

Category:

Engineering

Download: 0

Comment: 0

1,004

views

Comments

Description

I present "Hand posture recognition base on bottom-up structured deep convolutional neural network with curriculum learning" at ICIP 2014.
Download ICIP2014 Presentation

Transcript

  • 1. 㻴㼍㼚㼐㻌㻼㼛㼟㼠㼡㼞㼑㻌㻾㼑㼏㼛㼓㼚㼕㼠㼕㼛㼚㻌㻮㼍㼟㼑㼐㻌㼛㼚㻌㻮㼛㼠㼠㼛㼙㻙㼡㼜㻌㻿㼠㼞㼡㼏㼠㼡㼞㼑㼐㻌㻰㼑㼑㼜㻌㻯㼛㼚㼢㼛㼘㼡㼠㼕㼛㼚㼍㼘㻌㻺㼑㼡㼞㼍㼘㻌㻺㼑㼠㼣㼛㼞㼗㻌䈊㼣㼕㼠㼔㻌㻯㼡㼞㼞㼕㼏㼡㼘㼡㼙㻌㻸㼑㼍㼞㼚㼕㼚㼓䈊Takayoshi Yamashita1, Taro Watasue21Chubu University, 2Tome RD
  • 2. 㻯㼍㼚㻌㼥㼛㼡㻌㼒㼕㼓㼡㼞㼑㻌㼛㼡㼠㻌㼠㼔㼕㼟㻫1x2 +1∫ dx
  • 3. 㻯㼍㼚㻌㼏㼔㼕㼘㼐㻌㼒㼕㼓㼡㼞㼑㻌㼛㼡㼠㻌㼠㼔㼕㼟㻫1x2 +1∫ dx
  • 4. 㼀㼛㻌㼟㼛㼘㼢㼑㻌㼠㼔㼕㼟㻌㼑㼝㼡㼍㼠㼕㼛㼚It requires fundamental knowledge of math that studies along thecurriculum (with other knowledge form different classes) 1x2 +1∫ dxarithmetic equationdifferentialsquare integrationrootpsychics ………..
  • 5. 㻷㼑㼥㻌㼕㼐㼑㼍㻌㼛㼒㻌㼠㼔㼕㼟㻌㼜㼞㼑㼟㼑㼚㼠㼍㼠㼕㼛㼚Inspired from human’s knowledge acquisition ! Train good feature representation usingthe curriculum learning! Transfer the knowledge (networks) fromheterogeneous task
  • 6. 㻰㼑㼑㼜㻌㻯㼛㼚㼢㻚㻌㻺㼑㼠㼟! Deep architecture which consists of convolutional, sampling and fullyconnection layers [LeCun 1998]! It has translation invariance of object! CNN+ ReLu, dropout, Normalization, etc [Krizhevsky 2012]! Recognize the category of 1000 classes! Top performance in Large Scale Visual Recognition Challenge 2012
  • 7. 㻯㼡㼞㼞㼕㼏㼡㼘㼡㼙㻌㻸㼑㼍㼞㼚㼕㼚㼓! Train while changing difficulty of training dataset䚷䚷(= similar with Bootstrap, but different…)x1x2x3xiy1y2h1h2yjy1y2y3hjinitial training with simple set (square size)update with complexity set (various aspect ratio)We propose the novel curriculum learning which updates thenetwork from the heterogeneous taskY. Bengio, J. Louradour, R. Collobert, J. Weston, “Curriculum Learning”, ICML2009.
  • 8. 㼜㼞㼛㼜㼛㼟㼑㼐㻌㼙㼑㼠㼔㼛㼐• Train good feature representation using curriculum learning• Transfer the knowledge from heterogeneous taskhand gesture recognitionmain idea of proposed method-We train the network with two curriculum-Two curriculums are heterogeneous
  • 9. 㻼㼞㼛㼜㼛㼟㼑㼐㻌㼙㼑㼠㼔㼛㼐䚷ࠥ㼠㼞㼍㼕㼚㼕㼚㼓㻔㻝㻕ࠥTrain the networks as segmentation taskConvolutional Layer Pooling Layer fully connection LayerConvolutional Layer Pooling LayerBinarization layerInput data : gray scale imageground truth : hand segmented image
  • 10. 㻼㼞㼛㼜㼛㼟㼑㼐㻌㼙㼑㼠㼔㼛㼐䚷ࠥ㼠㼞㼍㼕㼚㼕㼚㼓㻔㻞㻕ࠥTransfer the networks to classification taskUtilize as initial parametersInput data : gray scale imageground truth : class labelupdating the parameters
  • 11. 㻼㼞㼛㼜㼛㼟㼑㼐㻌㼙㼑㼠㼔㼛㼐䚷ࠥ㼜㼞㼑㼐㼕㼏㼠ࠥClassify the object using only updated networks5Input data : gray scale imageoutput : class label
  • 12. 㻱㼤㼜㼑㼞㼕㼙㼑㼚㼠㻌䠄䠍䠅! Evaluation data! 㻢㻌㼏㼘㼍㼟㼟㼑㼟㻌㻦㻌㼔㼍㼚㼐㻌㼟㼔㼍㼜㼑㻌㼜㼛㼟㼑! 㻏㻌㼛㼒㻌㼠㼞㼍㼕㼚㼕㼚㼓㻌㼕㼙㼍㼓㼑㼟㻌㻦㻌㻞㻜㻷㻌㼕㼙㼍㼓㼑㼟㻌㼎㼥㻌㼐㼍㼠㼍㻌㼍㼡㼓㼙㼑㼚㼠㼍㼠㼕㼛㼚! 㻏㻌㼛㼒㻌㼠㼑㼟㼠㼕㼚㼓㻌㼕㼙㼍㼓㼑㼟㻌㻦㻝㻢㻜㻜㻌㻔㻌㼑㼍㼏㼔㻌㼏㼘㼍㼟㼟㻕! Comparison! 㼔㼍㼚㼐㻌㼟㼔㼍㼜㼑㻌㼏㼘㼍㼟㼟㼕㼒㼕㼏㼍㼠㼕㼛㼚! 㻯㼛㼙㼜㼍㼞㼕㼟㼛㼚㻌㼙㼑㼠㼔㼛㼐㼟䠖䚷䚷㻙㼀㼞㼍㼕㼚㼕㼚㼓㻌㼣㼕㼠㼔㼛㼡㼠㻌㼏㼡㼞㼞㼕㼏㼡㼘㼡㼙㻌㼘㼑㼍㼞㼚㼕㼚㼓䚷䚷㻙㼀㼞㼍㼕㼚㼕㼚㼓㻌㼣㼕㼠㼔㻌㼜㼞㼛㼜㼛㼟㼑㼐㻌㼏㼡㼞㼞㼕㼏㼡㼘㼡㼙㻌㼘㼑㼍㼞㼚㼕㼚㼓䚷䚷䚷
  • 13. 㻱㼤㼜㼑㼞㼕㼙㼑㼚㼠䠄䠎䠅! Network architecturelayer settinginput 䚷䚷䚷 input layer 40x40 pixel (gray scale image)1st convolutional layer kernel size䠖5x5# of kernel䠖32activation function䠖Maxout䚷2nd pooling layer pooling䠖max poolingsize 䠖2x23rd convolutional layer kernel size䠖5x5# of kernel䠖32activation function䠖Maxout䚷4th pooling layer pooling䠖max poolingsize 䠖2x25th fully connection layer # of nodes䠖200activation function䠖sigmoidoutput classification layer(binarization layer)# of nodes :6䚷䚷䠄or 1600 when segmentation task䠅
  • 14. 㻱㼤㼜㼑㼞㼕㼙㼑㼚㼠䠄䠏䠅! Training parameters! 㻏㻌㼛㼒㻌㼡㼜㼐㼍㼠㼑㼟䠖㻞㻜㻜㻷! 㼘㼑㼍㼞㼚㼕㼚㼓㻌㼞㼍㼠㼑䃔䠖㻜㻚㻜㻜㻡ࠥ㻜㻚㻜㻜㻣! 㼙㼕㼚㼕㻌㼎㼍㼠㼏㼔㻌㼟㼕㼦㼑䠖㻝㻜! 㼐㼞㼛㼜㼛㼡㼠㻌㻦㻌㻡㻜㻑䚷䚷䚷
  • 15. 㼀㼞㼍㼕㼚㼕㼚㼓㻌㼑㼞㼞㼛㼞㻌 ) ( '
  • 16. * %
  • 17. *
  • 18. * '* %* * ' % $ #! % !!%
  • 19. 㻼㼑㼞㼒㼛㼞㼙㼍㼚㼏㼑
  • 20. Ground Truth class
  • 21. without curriculum learning with curriculum learningclassification classGround Truth classclassification class
  • 22. 㼂㼕㼟㼡㼍㼘㼕㼦㼍㼠㼕㼛㼚㻌㼛㼒㻌㼗㼑㼞㼚㼑㼘㼟1stconvolutionallayer2ndconvolutionallayerwithout Curriculum learning with Curriculum learningtotal updating time : 200000 total updating time : 200000(segmentation: 50000 +recognition:15000)
  • 23. 㻵㼚㼠㼑㼞㼙㼑㼐㼕㼍㼠㼑㻌㼜㼍㼞㼍㼙㼑㼠㼑㼞㼟㻌㼢㼕㼟㼡㼍㼘㼕㼦㼍㼠㼕㼛㼚! updating time : 0 - 50000
  • 24. ! final parameters of binarization layer㻝㻥x1x2x3xiy1y2Yj㻵㼚㼠㼑㼞㼙㼑㼐㼕㼍㼠㼑㻌㼜㼍㼞㼍㼙㼑㼠㼑㼞㼟㻌㼢㼕㼟㼡㼍㼘㼕㼦㼍㼠㼕㼛㼚
  • 25. ! final parameters of binarization layer㻞㻜x1x2x3xiy1y2Yj㻵㼚㼠㼑㼞㼙㼑㼐㼕㼍㼠㼑㻌㼜㼍㼞㼍㼙㼑㼠㼑㼞㼟㻌㼢㼕㼟㼡㼍㼘㼕㼦㼍㼠㼕㼛㼚
  • 26. ! final parameters of binarization layer㻞㻝x1x2x3xiy1y2Yj㻵㼚㼠㼑㼞㼙㼑㼐㼕㼍㼠㼑㻌㼜㼍㼞㼍㼙㼑㼠㼑㼞㼟㻌㼢㼕㼟㼡㼍㼘㼕㼦㼍㼠㼕㼛㼚
  • 27. ! final parameters of binarization layer㻞㻞x1x2x3xiy1y2Yj㻵㼚㼠㼑㼞㼙㼑㼐㼕㼍㼠㼑㻌㼜㼍㼞㼍㼙㼑㼠㼑㼞㼟㻌㼢㼕㼟㼡㼍㼘㼕㼦㼍㼠㼕㼛㼚
  • 28. 㼔㼍㼚㼐㻌㼟㼔㼍㼜㼑㻌㼟㼑㼓㼙㼑㼚㼠㼍㼠㼕㼛㼚! Extract hand region from gray scale image inclutter background
  • 29. 㻯㼛㼚㼏㼘㼡㼟㼕㼛㼚! We propose the training method of Deep ConvolutionalNeural Networks with curriculum learning! As the curriculum, the method transfer the network fromheterogeneous task (segmentation = classification)! The method is able to improve the feature representation! Future works䚷䚷apply to other objects and new curriculum
  • 30. Thank you for your attention
  • Fly UP