Download presentation
Presentation is loading. Please wait.
Published byMolly Gibson Modified over 8 years ago
1
Facial Detection via Convolutional Neural Network Nathan Schneider
2
Introduction Why do this project? Implements CNN Important applications Hard problem It would be cool! AT&T Database 400 92x112 images
3
Caffe Matlab Visual Studio
4
Next Idea – Make it myself Why? Learn more Verify understanding Remember better More fun!
5
C Why C? I already know it Matlab is very specialized C is powerful and quick
6
Basic review Convolutional Layer Runs filter over every part of input image Max-pooling Takes max over small region Reduces complexity 3D Neurons 4D weights!
7
Structure Convolutional Layer (4) Max Pooling (2x2) Hidden Layer Output Layer
8
Challenges Parameters Currently 21 parameters determining CNN Only 3 for input and 1 for output => 17 variable parameters Scale differences Deltas quarter every layer Tried gradient approach 100k, 2.5k, 400, 70, 2.5
9
Challenges Debugging Over 1,000 lines of code for CNN Mostly for loops LOTS of for loops Impossible to step through 4D weights lots of values to check
11
Challenges Max Pooling Sometimes erases important data Negative weights How to erase background noise and keep important variations? Deviation from average!
12
Challenges AT&T Data set format PGM – Portable Gray Map How to interface with images? Netpbm Library issues
13
Results Average of 70% classification after 100 epochs with simple example Actual data gave constant positive
14
Future Improve speed (2-3 min per epoch) GPU utilization? Tweak parameters User friendly interface / any interface
15
References AT&T Database http://www.cl.cam.ac.uk/research/dtg/attarchive/facedat abase.html http://www.cl.cam.ac.uk/research/dtg/attarchive/facedat abase.html Intro to CNN http://neuralnetworksanddeeplearning.com/chap6.html Netpbm http://netpbm.sourceforge.net/doc/index.html
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.