130 likes | 360 Views
Eye tracking to enhance facial recognition algorithms. Balu Ramamurthy Brian Lewis December 15, 2011. Introduction. Facial recognition is growing security concern Best recognition algorithm is human brain Wanted to find a way to use brain information in recognition
E N D
Eye tracking to enhance facial recognition algorithms Balu Ramamurthy Brian Lewis December 15, 2011
Introduction • Facial recognition is growing security concern • Best recognition algorithm is human brain • Wanted to find a way to use brain information in recognition • If we identify areas humans use to recognize faces, we can get unique results in algorithms
Contents • Biometrics Background • Eye Tracking Experiment • Facial Recognition Experiment • Facial Recognition Results • Conclusion • Future Work
Biometrics Background • 2 types of biometrics, identification and verification • Verification consists of confirming an identity • Identity comes from selecting correct person from a group of candidates • Current algorithms use features extracted from images
Eye Tracking Experiment • Used 10 males and 10 females • Ran identification and verification experiments • Females much better at identifying faces • Conducted identification and verification experiments
Verification Experiment • 2 Normalized faces shown to participant • Participant asked to say if same person or different person
Identification Experiment • Participant looks at image of face for as long as needed. • Then shown 2 by 3 grid of normalized faces to identify correct face
Facial Recognition Procedure • Each correct image broken up in to 7 by 7 grid • Percentage of fixations for each block extracted.
Facial Recognition Experiment • Experiment 1 gave each block equal distribution • Experiment 2 blocks weighted 0-3 with equal number of blocks in each weight • Experiment 3 blocks given weights of 0-4 based on fixation percentages • Experiment 4 only blocks of 100% fixation were used in algorithms
Conclusion • No significant recognition rate improvement • Blocks with 100% fixation account for 50% of accuracy • Trial and error in experiments 3 and 4 give hope for future work
Future Work • Develop algorithm to properly weight boxes • Look at using new tasks for eye tracking • Try new facial recognition algorithms on data • Run experiments using specific facial regions