A Framework for PreMaturation AgeProgressed Face Recognition

1
A Framework for Pre-Maturation Age-Progressed
Face Recognition

• Karl Ricanek, Jr. Ph.D.

2
Overview

• Research Background
• Current Research Efforts
• Past Areas of Research
• Problem Statement
• Matching Algorithm
• Feature Set
• Pre-Maturation FR
• Closing
• Follow On Work
• Q A

3
Research

• Current Research
• Craniofacial Morphological Changes 3D Models of
  the Human Face Head
• Team Eric Patterson (CSC/UNCW) Midori Albert
  (ANT/UNCW)
• Funding Funding by DoD
• 3D Face Recognition Algorithms
• Team Winser Alexander (EE/NC State) Jung Kim
  (EE/NC AT)
• Funding Developing proposal
• Cognitive Models for Person Identification
• Team Dale Cohen (PSY/UNCW)
• Funding Still forming group. No funding sought.

4
Research

• Student
• Object Tracking Using Color Information
• Student Patrick Cash (Senior)
• Image Restoration and Processing
• Student Robert Harrison (Junior/DIS)
• Web Design/Development
• Student Brandon Bullock (Freshman/CTE funded)

5
Research

• Past Areas of Research
• Artificial Intelligence
• Expert Systems
• Machine Learning
• Neural Network, Fuzzy Systems, Genetic Algorithms
• Image Processing
• Compression, Object Recognition, Restoration
• Signal Processing
• Classification, Analysis
• Optics
• Interferometry, metrology, machine vision,
  birefringence, polarization, polarization mode
  dispersion

6
Problem Statement

• The development of a robust face recognition
  system for pre-maturation craniofacial growth
  changes.
• Large changes in the head and face over small
  periods of time occur during the maturation
  process.
• Image based techniques (Eigenface, fisherfce,
  etc.) are at a disadvantage
• Feature based techniques may be morphed
  (projected).
• Automatic face recognition for children has not
  been an area of research.

7
Matching Algorithm

• Annealed Hopfield Neural Network
• Hopfield Neural Network
• fully connected recurrent network
• converges to local minimum
• asynchronous updates
• Mean Field Annealing
• deterministic annealing
• decreases computation overhead
• optimal solution ensured under conditions

8
Hopfield NN

• Optimization with Hopfield Structure
• 1. Develop objective (cost) function
• 2. Encode problem into a set of variables Vi, 0lt
  Vi lt 1, neuromorphic form
• 3. Derive the energy function, E(V)
• 4. Determine the function of the neuron output
• 5. Initialize network
• 6. Iterate until stopping condition
• 7. Decode the results

9
Hopfield NN

• HNN Convergence

Local convergence is ensured. What about the
global solution?
10
Annealing

• Simulated Annealing
• Optimization model based on the physical process
  of annealing.
• Randomly, perturbs the system towards its thermal
  equilibrium.
• Global optimal ensured if its corresponding Tc is
  in the cooling schedule.

11
Annealed Hopfield Network

• Mean Field Annealing Energy
• Hopfield Energy

12
Graph Matching

• Graphs (probe and model) are defined by
• local features distance, and
• relational features angles.
• Compatibility measure determines match between
  probe and model nodes

13
Graph Matching with Hopfield

• Develop an appropriate representation of the
  graph matching problem into the Hopfield energy
• create objective function
• encode problem into neuromorphic form
• derive energy function

14
Graph Matching with Hopfield Cont.
15
Graph Matching with Hopfield Cont.
16
Graph Matching with Hopfield Cont.
17
Graph Matching with Hopfield Cont.
18
Feature Set

• A feature set should be
• consistent and reproducible,
• a unique discriminator, and
• have compact encoding.
• Local features are
• in-plane rotation invariant,
• translation invariant, and
• can be scaled normalized.

19
Feature Set of a Digital Face

• Analysis of a digital image of a face under
  various pose articulation was performed.
• Twenty seven anthropometric landmarks (AL) were
  identified by this research for pose varying
  digital faces.

20
Feature Set Identified
21
Feature Set Illustrated
AL Features Frontal
AL Features Profile
22
Feature Set

• Autocorrelation Discrimination Metric
• Discrimination quantity was formulate based on a
  weighted measure of positive node matches, false
  node matches, and nodes not matched.
• Confidence W1 x Correct W2 x False W3 x
  Unavailable
• C 1, all nodes of probe match all nodes of
  gallery
• Each subject was compared to the gallery over
  view.
• Autocorrelation Discrimination Results
• Average over all (13) views ranged from 0.211 to
  0.267
• Each model perfectly matched, C1.0, their
  corresponding gallery image over all views, no
  other perfect match occurred.

23
AHNN Results

• Probe set
• Scale variation and face gesture variation
• Recognition Rate
• No false positive matches were made for frontal
  view 99.99
• Robust under pose view matching 90

24
AHNN Illustration
25
Pre-Maturation FR

• Extensive literature search has failed to find
  research in this area outside of Forensic Art.
• Some areas of interest for this work
• missing children,
• reconstructive forensics, and
• security.

26
Pre-Maturation FR

• During the maturation period (birth to
  post-adolescence) rapid growth is experienced.
• The degree of craniofacial growth is dependent
  on
• age,
• gender, and
• ethnicity.
• Craniofacial growth can be measured by the
  anthropometric landmarks identified by this work.

27
Pre-Maturation FR

• Procedures for age projection
• Take measurements based on AL (photogrammetry).
• Compare measurements to population norms for age
  to establish a baseline
• Project measurements to population norms for the
  new age, keeping measurement positions within
  population from baseline

28
Experimental Results

• Results are not available at this time.
• Creation of database not complete.

29
Closing

• Follow on work
• Complete evaluation database
• Test AHNN FR on database
• Characterize the effects of ethnicity
• Characterize admixture (multi-ethnic)
• Questions Answer