age estimation: a classification problem
DESCRIPTION
AGE ESTIMATION: A CLASSIFICATION PROBLEM. HANDE ALEMDAR, BERNA ALTINEL, NEŞE ALYÜZ, SERHAN DANİŞ. Project Overview. Subset Overview. Aging Subset of Bosphorus Database: 1-4 neutral and frontal 2D images of subjects 105 subjects Total of 298 scans Age range: [18-60] - PowerPoint PPT PresentationTRANSCRIPT
AGE ESTIMATION: A CLASSIFICATION PROBLEM
HANDE ALEMDAR, BERNA ALTINEL, NEŞE ALYÜZ, SERHAN DANİŞ
Project Overview
Subset Overview
• Aging Subset of Bosphorus Database:– 1-4 neutral and frontal 2D images of subjects– 105 subjects– Total of 298 scans– Age range: [18-60]– Age distribution non uniform: average = 29.9
Project Overview
• Aging images of individuals is not present• Aim: Age Estimation based on Age Classes• 3 Classes:
• Age<26 -> 96 samples• 26 <= Age <= 35 -> 161 samples• Age>36 -> 41 samples
Preprocessing
• Registration• Cropping• Histogram Equalization• Resizing
SUBSPACE ANALYSIS FOR AGE ESTIMATION
Neşe Alyüz
Age Manifold
• Instead of learning a subject-specific aging pattern, a common aging trend can be learned
• Manifold embedding technique to learn the low-dimensional aging trend.
Image space: Labels:Low-dim. representation:
d<<D
Mapping:
Orthogonal Locality Preserving Projections - OLPP
• Subspace learning technique• Produces orthogonal basis functions on LPP• LPP:
The essential manifold structure preserved by measuring local neighborhood distances
• OLPP vs. PCA for age manifold: OLPP is supervised, PCA is unsupervised OLPP better, since age labeling is used for learningX Size of training data for OLPP should be LARGE enough
Locality Preserving Projection - LPP
• aka: Laplacianface Approach
• Linear dimensionality reduction algorithm
• Builds a graph: based on neighborhood information
• Obtains a linear transformation:Neighborhood information is preserved
LPP• S: similarity matrix defined on data points (weights)• L = D – S : graph Laplacian• D: diagonal sum matrix of S
measures local density around a sample point• Minimization problem:
with the constraint :
=> Minimizing this function: ensure that if xi and xj are close then their projections yi and yj are also close
LPP
• Generalized eigenvalue problem:
• Basis functions are the eigenvectors of:
Not symmetric, therefore the basis functions are not orthogonal
OLPP
• In LPP, basis functions are nonorthogonal– > reconstruction is difficult
• OLPP produces orthogonal basis functions– > has more locality preserving power
OLPP – Algorithmic Outline
(1) Preprocessing: PCA projection(2) Constructing the Adjacency Graph(3) Choosing the Locality Weights(4) Computing the Orthogonal Basis Functions(5) OLPP Embedding
(1) Preprocessing: PCA Prjection
• XDXT can be singular• To overcome the singularity problem -> PCA• Throwing away components, whose
corresponding eigenvalues are zero.• Transformation matrix: WPCA
• Extracted features become statistically uncorrelated
(2) Constructing The Adjacency Graph
• G: a graph with n nodes
• If face images xi and xj are connected (has the same label) then an edge exists in-between.
(3) Choosing the Locality Weights
• S: weight matrix• If node i and j are connected:
• Weights: heat kernel function• Models the local structure of the manifold
(4) Computing the Orthogonal Basis Functions
• D: diagonal matrix, column sum of S• L : laplacian matrix, L = D – S• Orthogonal basis vectors: • Two extra matrices defined:
• Computing the basis vectors:– Compute a1 : eigenvector of with the greatest eigenvalue – Compute ak : eigenvector of
with the greatest eigenvalue
(5) OLPP Embedding
• Let:
• Overall embedding:
Subspace Methods: PCA vs. OLPP
• Face Recognition Results on ORL
Subspace Methods: PCA vs. OLPP• Face Recognition Results on Aging Subset of the Bosphorus
Database
• Age Estimation (Classification) Results on Aging Subset of the Bosphorus Database
FEATURE EXTRACTION: LOCAL BINARY PATTERNS
Hande Alemdar
Feature Extraction• LBP - Local Binary Patterns
Local Binary Patterns• More formally
• For 3x3 neighborhood we have 256 patterns• Feature vector size = 256
where
Uniform LBP• Uniform patterns can be used to reduce the length of
the feature vector and implement a simple rotation-invariant descriptor
• If the binary pattern contains at most two bitwise transitions from 0 to 1 or vice versa when the bit pattern is traversed circularly Uniform– 01110000 is uniform– 00111000 (2 transitions)– 00011100 (2 transitions)
• For 3x3 neighborhood we have 58 uniform patterns• Feature vector size = 59
FEATURE EXTRACTION: GABOR FILTERING
Serhan Daniş
Gabor Filter
Band-pass filters used for feature extraction, texture analysis and stereo disparity estimation. Can be designed for a number of dilations and rotations.
The filters with various dilations and rotations are convolved with the signal, resulting in a so-called Gabor space. This process is closely related to processes in the primary visual cortex.
Gabor Filter
A set of Gabor filters with different frequencies and orientations may be helpful for extracting useful features from an image. We used 6 different rotations and 4 different scales on 16 overlapping patches of the images. We generate 768 features for each image.
CLASSIFICATIONBerna Altınel
EXPERIMENTAL DATASETS
1. FEATURES_50_45(LBP) 2. FEATURES_100_90(LBP)3. FEATURES_ORIG(LBP)4. FEATURES_50_45(GABOR)5. FEATURES_100_90 (GABOR)
Estimate age, just based on the average value of the training set
Experiment #1
EXPERIMENTAL RESULTS:INPUT METHOD Mean(MRE) % Number of
Correct classifications
Number of missclassifications
Features_50_45-LBP
Estimating the Average Age 17.31 162 / 298 135 / 298
Features_100_90-LBPFeatures_orig-LBP
Features_50_45(GABOR)Features_100_90(GABOR)
K-NEAREST-NEİGHBOR ALGORİTHM
Experiments #2
The K-nearest-neighbor (KNN) algorithm measures the distance between a query scenario and a set of scenarios in the data set.
EXPERIMENTAL RESULTS:
INPUT Features_50_45(LBP)
Features_100_90(LBP)
Features_orig(LBP)
Features_50_45(GABOR)
Features_100_90(GABOR)METHOD
kNN-1(Euc Dist)
MRE(%):5.05 MRE(%):4.14 MRE(%):11.11
MRE(%):3.88 MRE(%):3.75
kNN-2(Euc Dist)
MRE(%):6.77 MRE(%):5.17 MRE(%):11.97
MRE(%):4.92 MRE(%):5.08
kNN-3(Euc Dist)
MRE(%):7.50 MRE(%):6.06 MRE(%):12.50
MRE(%):5.79 MRE(%):5.96
kNN-5(Euc Dist)
MRE(%):10.40
MRE(%):9.36 MRE(%):13.15
MRE(%):11.02
MRE(%):10.93
kNN-10(Euc Dist)
MRE(%):12.34
MRE(%):11.57
MRE(%):14.16
MRE(%):13.86
MRE(%):14.13
kNN-15(Euc Dist)
MRE(%):12.85
MRE(%):12.30
MRE(%):14.35
MRE(%):14.57
MRE(%):14.98
Features_50_45(LBP)
Features_100_90(LBP)
Features_orig(LBP)
Features_50_45(GABOR)
Features_100_90(GABOR)
METHOD
Average MRE(%):15.72
MRE(%):15.01
MRE(%):15.58
MRE(%):16.31
MRE(%):16.55
MissClass:35 / 298
MissClass:32 / 298
MissClass:31 / 298
MissClass:32 / 298
MissClass:27 / 298
CorrectClass:262 / 298
CorrectClass:265 / 298
CorrectClass:266 / 298
CorrectClass:265 / 298
CorrectClass:270 / 298
IN PROGRESS:1. Parametric Classification
2. Mahalanobis distance can be used as the distance measure in kNN.
[2 [2
1. Other distance functions can be analyzed for kNN:
2. Normalization can be applied:
POSSIBLE FUTURE WORK ITEMS: