face recognition and biometric systems

Face normalisation

Face Recognition and Biometric Systems

Plan of the lecturePlan of the lecture

Normalisation – task definitiontesting issuestesting issues

Geometric normalisationLighting normalisationAdvanced normalisation issuesAdvanced normalisation issues


Face recognition processFace recognition process

Detection Normalisation

FeatureFeature vectors Feature extraction

Feature vectorscomparison


Normalisation generalNormalisation – general

Image preparation for feature extraction

similar properties of generated imagesgeometryconditions (e.g.: lighting, expression)occlusions

Intra-class differences minimisedExtra-class differences notExtra class differences not influenced



Effectiveness criteriavisual effectrecognition performanceg p

Detection error influences normalisation resultnormalisation result



Perfect detectionreal location of face and facial featuresdata input by humanp y

Elimination of detection error propagationpropagationBetter assessment of subsequent recognition stages


Geometric normalisationGeometric normalisationR i tRequirements:

constant image sizefixed eye positionsfixed eye positionsfrontal orientation (soft requirement)

Frontal faces – goal:given positions of eyesaffine transform

Actions:Actions:clippingrotationscaling

Time for example


Geometric normalisationGeometric normalisationS d ti i tiSpeed optimisation

larger image = more time consumed

Optimal algorithm:Calculate rotation angle1. Calculate rotation angle

2. Find and clip the ROI3. Rotate the clipped image4. Clip againp g5. Scale to the defined size


Laboratory reference (ex 2)Laboratory reference (ex 2)

Function parametersE itiEye positions:

left (49, 24)right (15, 24)

IPP referenceIPP referenceRotateCenterResizeResize

Operations...Face Recognition and Biometric Systems

Lighting normalisationLighting normalisation

Lighting codnitions affect effectivenessNormalisation techniques:Normalisation techniques:

global filteringlocal modificationslocal modifications

Histogram modifications:stretchingstretchingequalisationfitting to the average face histogramfitting to the average face histogram

Filtering



Average faceAverage face

∑M

iiM 1

1= xμ=iM 1

M – number of faces in a setx – a single face vectorg



With histogram fitting:

Without histogram fitting:



Brightening filters – example of effects



Directional lighting:strong influence on the imagerecognition effectiveness much worseg

Light direction normalisation:li ht l d t tilight angle detectioncompensation to the frontal light conditions



Mirror reflectionLighting compensation masks:

lighting symmetrisationlighting symmetrisationcompensation to the average

d l b d kmodel-based mask

Filtering based on lighting modelg g gCompensation based on lighting model



Mirror reflectionCondition: no information in one image halfI h lfImage half recoveryApplicable to frontal faces onlyfaces onlyBrightness and angle thresholdingthresholding


Li hti li ti kLighting normalisation - masks

Image-based lighting compensation masksd k li ht ddark areas lightenedhighlights darkened

k h lMask imposition on the original image:additionmultiplicationadvanced imposition – to be investigated...


Li hti li ti k

S t i k

Lighting normalisation - masks

Symmetric mask

Compensation to the averageCompensation to the average



Lighting compensation – face modelDetection of lighting direction

based on average 3D face modelbased on average 3D face modelclassifiers (SVM, PCA)

C ti b d 3D d lCompensation based on 3D modelmask generationg

Works correctly for artificial data



Lighting-model based compensationLighting model based compensationInitial object (ambient light):

[ ]a[m,n]

Illuminated object:jc[m,n] = a[m,n] • I[m,n]



Light – low frequencies in the imageg q gLow frequencies elimination:

ln(c[m n]) ln(I[m n]) + ln(a[m n])ln(c[m,n]) = ln(I[m,n]) + ln(a[m,n])HP{ln(c[m,n])} ≈ ln(a[m,n])a’[m,n] = exp{HP{ln(c[m,n])}

Theory seems niceTheory seems nice...


Advanced normalisationAdvanced normalisation

Head rotation normalisationf t l i d i dfrontal image desired

Face expression normalisationneutral expressionexpression detectionp

Elimination of occlusionsglassesglassesbeard and moustache


Non frontal imagesNon-frontal images

Significant influence on recognition effectivenessNormalisation (rotation):

3D2D + depth map

The most serious problem: angle detectionFace Recognition and Biometric Systems

p g

SummarySummary

Normalisation – important step in face itirecognition process

Tasks:as ssize and position normalisationimage properties normalisationimage properties normalisation

Many areas for further research


Thank you for your attention!Thank you for your attention!


face recognition and biometric systems

Documents