Palm Biometrics Flashcards
(19 cards)
D. Zhang Palm Recognition Method
Input Palm Into Machine
Illuminate with ring light
Capture with Lens
A/D Converter
CCD Camera
D Zhang Preprocessing
Resize Original Hand Image (384x284)
Binarize Image
Boundary tracking, isolate key points and build coordinate system.
Extract central part by filtering extraneous data
Handprint Principle Lines
Principle Lines cannot be used for recognition as different people have can have similar principle lines.
Low Resolution Handprints
Resolution is usually low so details such as wrinkles are not clear
Incorrect Hand Placement
To avoid extracting features from background due to incorrect hand placement, palm masks are used.
Hand Placement Gabor Wavelets
Features obtained from 2-D Quadrature Gabor Wavelets:
- First Binary bit code (real part)
- Second Binary Bit Code (Imaginary Part)
- Corresponding Masks
Palm Matching
Palm matching is performed using normalised Hamming distance based on real part, imaginary part and mask.
Palm Verification
Imposter Palms hover around 0.33 to 0.5 HD as expected, Genuine palms more spread HD than iris but spread from 0 to 0.33.
Clear threshold between the two.
Receiver Operator Characteristic
Equal Error Rate 0.6%, good but not perfect, comparable with other palmprint approaches.
Effect of time on handprints
Time has little effect and palm does not change for years.
HD were smallest between palms from different times.
Effect of illumination on handprints
Handprint recognition is robust to illumination.
HD were again smallest between different illumination levels.
Eigen palms
Eigen palms contained far more database samples than previous approaches.
Exhibited a very high recognition rate.
Palm Biometric Extraction Modality 1
Acquire Hand Image
Binarize hand image and find best fitting ellipse using circular Hough Transform
Rotate handprint according to orientation of fitted ellipse
Distance transform of the handprint assigns distance from each pixel to boundary of the hand, max pixel (most dense point) is considered centre
Estimate palmprint region around estimated centre point.
Reverse the rotation.
Palm Biometric Extraction Modality 2
Same as previous. However, greyscale after rotating.
Use Region of Interest (ROI) extracted from centre. Use erosion to refine boundaries.
Palmprint Feature Extraction
Segmented Image is normalised
Image is filtered using a direction mask at orientations (0,90,45,135)
Max of these 4 at each pixel are taken to combine to G
Features are the standard deviation of blocks
Multimodal Fusion
Combining both hand geometry and palm print consistently performs better
Palmprint Verification (Kumar et al.)
Cosine similarity is computed between two feature vectors
Partly Model Based approach to hand recognition
Models palm by regions, principal lines and datum points.
Laws filters convolve palm images.
Statistical Features are taken as features.
