Face and Fingerprint Flashcards
(28 cards)
Optical vs Capacitive Fingerprint Sensors
Optical use a diffused light source and a sensor which captures light bounced off fingerprint.
Capacitive generates charge from touching sensor in different locations.
Fingerprint Patterns
Set of ridges (sweat bands)
Can be arch, loop or whorl.
Minutiae are the terminations (end points) or the bifurcations (splitting points)
Pattern depends on resolution.
Basic Fingerprint Pre-processing
Histogram equalisation: increased contrast in the image, making print pattern more visible .
Median Filtering: Makes edges more prominent
Thresholding: works well, transforms into binary image.
Fingerprint Enhancement
Normalisation: normalise to specific mean and variance
Orientation Estimation: Using the angle of the gradient of the image
Local Frequency is estimated
Compose mask for unrecoverable regions.
A bank of Gabor filters tuned to a local ridge orientation and ridge frequency are applied.
Fingerprint Recognition Approaches
Minutiae - choose maximum alignment of minutiae pairings.
Correlation/Distance - maximise match between fingerprint images.
Ridges - maximise match of selected ridge features, such as local orientation, frequency, shape and texture.
Minutiae Detection and Matching
Termination and Bifurcation. Also can have false minutiae caused by noise ridgelines.
Can be manual or automatic.
Define feature vector based on Minutiae position, as well as distances and angles between them.
Texture Based Matching
Locate Reference Point by correlating fingerprint with template and finding maximum.
Divide image inter sectors, and normalise each sector.
Apply Gabor Filters to extract texture information.
Compute Average Absolute Derivation to generate finger code.
Calculate Euclidean distance between template finger code.
Match result.
Fingerprint Challenges
Pressure/Skin Deformation
Image quality and forensic use
Skin condition
Legal issues, is fingerprint unique and permanent
Spoofing
Pressure Correction
Finger pressure against sensor is not uniform, but decreases moving from the centre.
Close contact region: Skin does not move
Transitional Region: Skin distorted
External Region: Skin moves
Can distort image to correct skin distortion.
Automated Face Recognition Approaches
Holistic - Image as a whole
Model Based - Recognition by parts
PCA
Reduce Data dimensionality, PCA reduces data to principle components which capture most variance in the data.
Method:
- Subtract the average image from every dataset image
- Compute the covariance matrix of the zero-mean dataset
- Find the eigen values and vectors from the covariance matrix
- Vectors = direction of variation. Values = amount of variation
- Order eigen vectors by eigen values
- Select top k eigen values for k principle components.
Eigenface Overview
Calculate input image weights (how much of each eigen face is this image)
Check if its even a face
If face, classify weight pattern as either known or unknown person
Can optionally update faces if unknown.
SVD
Singular Value Decomposition. Matrix factorisation can be used to compute principal components without explicitly executing the PCA algorithm.
Eigenface Approach
Collect Image Dataset (eg. four images per 10 people, m=40)
Calculate 40x40 matrix, find eigen vectors and values, choose top M’ values to represent face features.
Combine the normalised training set of images to produce M’ eigenfaces
For known individuals, calculate the class vector by averaging eigenface patterns across subjects. Choose thresholds
For new faces, calculate pattern vector and distances, if lower than threshold then match. If match then add to dataset.
Eigenface Results
Performance declined as the following varied:
- lighting
- head size
- orientation
- size
- combinations of the above
Haar Wavelets
Simple black and white patterns used to detect specific features in an image, like distinct facial regions.
Feret Database
Face image dataset containing variations in facial expression, lighting, camera and time.
Includes frontal images, duplicates taken at different times in the year and pose variations.
More Face databases
Notre Darme
- unstructured lighting condition recorded in hallway
Texas
- The University of Texas Video Database includes videos of people showing facial speech, laughter, and disgust.
Face Recognition in changing Illumination
Recognition Rate increased after pre-processing:
- Illumination normalisation using gamma correction
- DoG filtering - difference of gaussians
- Masking
- Equalisation of Variation
Varying Gamma
Increase brightness in darker parts of images. Can locally calculate and increase gamma to achieve a unity gamma.
Local Binary Patterns
Simple texture descriptor to capture local spatial patterns
- chooses 3x3 window
- takes centre pixel
- if neighbour pixel > centre then 1 in output code
- produces 8 bit code
X. Tan and B. Triggs Approach
The system starts with a raw image, applies pre-processing, and extracts multiple types of features (like Gabor, LBP, and others). Each feature set is passed through a kernel-based matcher to compute similarity scores. These scores are then normalized using Z-score normalization and finally combined (score fusion) to improve overall recognition performance.
Challenges in Face Recognition
Lighting
Viewpoint
Occlusion
Resolution
Facial Expression
Ageing
Make-up
Facial Expression Approach Analysis
Gabor Wavelets: Help detect fine grained muscle movements
Component Based: calculate difference between action and neutral faces.
- ICA (independent component analysis) captures muscle patterns
- PCA captures main facial expression variation
