Image Recognition Christian Cosgrove Kelly LiRebecca Lin Shree NadkarniSamanvit Vijapur Priscilla Wong Yanjun YangKate YuanDaniel Zheng Drew University New Jersey Governor’s School in the Sciences
Dog
What is image recognition?
Machine Learning Testing and Training
Database Feature Extraction Classification
Digit Database 515 images per digit
AT&T Face Database 40 subjects 10 images per subject
NJGSS Face Database 40 subjects 10 images per subject
Feature Extraction Vertical Edges Intensity Horizontal Edges Within each patch:
Classification Nearest NeighborCentroid
Classification Nearest NeighborCentroid
Output: Confusion Matrix
Network Graph
Digit Classification Performance WHAT IS “N PATCHES”? Number of patches Nearest Neighbor Centroids Number of patches Percent Accuracy
Facial Classification Performance Nearest Neighbor Centroids Number of patches Percent Accuracy
Colored Nearest Neighbor Colored Centroids Grayscale Nearest Neighbor Grayscale Centroids Number of Training Pictures Percent Accuracy
Conclusions Optimal Performance Digits: 588 features, 14x14 patches 91% AT&T Grayscale Faces: 480 features, 16x8/8x4 patches 99% ± 0.67% NJGSS RGB Faces: 360 features, 16x8/8x4 patches 96% ± 1.20%
Conclusions Improvements Prioritization of feature More training Tolerance for rotation and reflections of subjects Tolerance for background differences/lighting Applications Personal security Social media Robotics Medical examination
Acknowledgements Dr. Minjoon Kouh Mr. Michael Clancy NJ Governor’s School in the Sciences Dr. Adam Cassano Dr. Steve Surace Yumi Kouh Bayer Health Care Dr. Robert Mayans AT&T Laura (NJGSS ’86) and John Overdeck NJGSS Alumnae and Parents of Alumnae Board of Overseers, New Jersey Governor’s Schools State of New Jersey Drew University And all of NJGSS’s generous sponsors!