Video Based Palmprint Recognition
Chhaya Methani and Anoop M. Namboodiri
Center for Visual Information TechnologyInternational Institute of Information Technology Hyderabad
http://cvit.iiit.ac.in
Camera Based Palmprint Recognition Unconstrained Palm Imaging Video Based Palmprint Recognition
• Towards civilian biometric applications Low resolution web cameras Unrestricted & unconstrained imaging
• Palmprint as a biometricUnique Easy to capture User ConvenienceHigh mobilityEasy integration simplifies use in a multibiometric setup
• Challenges Background Contrast Noise Illumination Pose & scale
• Previous Work
Hand extraction from background has been attempted [1] Pixel noise is dealt by selection of good features viz. Gabor[2]
Noisy artefacts like blur, low image quality, shadow effects etc Illumination variations addressed by Intensity Normalization[2]
Not effective for low contrast or fully saturated pixels Scale variations[1]
Only when image is parallel to imaging plane Pose variations[3]
Accuracies can be improved by imbibing illumination invariance
• Pose and illumination Invariance: Two fold challenge
Matching samples from the same user under different lightingDetect and discard completely washed out samples
• Change in pose causes difference in reflection on different parts of the hand
Can be used to advantage by using more frames with different lighting• Video, instead of a single image, can capture more variations on the user’s palm
The user is asked to move the palm a little while recording the video to capture lighting variations
• Completely washed out samples are rejected by measuring texture content of the image, viz mean & variance of edge response
Aplplications of Camera Based Palmprint Imaging a) Access Control ,b) Multibiometric Image Capture c)Mobile unlocking
[1] JDoublet, et al. “Contactless hand Recognition Using Shape and Texture Features”, ICSP 2006
[2] Zhang et al. “Online Palmprint Recognition”, PAMI 2003
User 1 User 2
Figure showing difference in two images taken from the same user
[3] Methani,Namboodiri. “Pose Invariant Palmprint Recognition Recognition”, ICB 2009
Hand extraction from background
Video Capture & Pre-Processing
• Video Capture Video of duration approx 2 sec is recorded Users are advised to hold the hand loosely and allow some motion of hand to instigate illumination variation with view variation The background and the camera is fixed, with no restrictions on pose lighting etc Few frames from a recorded video
• Preprocessing Detect unique frames based on camera capture rate using background substraction Correct in-plane rotation & extract palm image Compute edge map Reject samples having very low texture; identified by the quality of the edge map obtained
Registration
• Important to overlay images separated in timeline before combining Registration usually done by finding corresponding landmark points in the two texturesThe transformations then found is used to align the two views In the absence of robust keypoints on palm, the entire edge map is used to iteratively come closer to the edge map of the consecutive frameSince the motion of hand is smooth, any two frames can be aligned by adding transformations of consecutive frames
Euclidean distance and gradient direction gives the nearest point in the consecutive frameThis is the approximate corresponding pointThe same process is followed for each point on the edge mapBased on all the point matches, a transformation is computed and matches improved iteratively Matching the entire pattern ensures a good matching
Results and Experiments
Observations and Conclusions
• Reason for average rule working better than max or second max seems to be the inexact registration • Slow rise in GAR(Genuine Accept Rate)initially in ROC curve indicates presence of some imposter scores before genuine scores
Indicates washed out samples• These samples are then rid of when thresholding is applied, saving the recognizer computational efficiency otherwise spent on “bad samples”
Conclusions and Future Work• Video Based Palmprint Recognition results in better accuracies than when single image is used due to the presence of extra information in the additional frames • EER reduced from 12.79% to 4.70% on frame combination and further to 1.9% on removing the washed out samples• As a follow up to this work, an improvement in registration process shall improve combination strategies and hence result in better strategies
a)Two images from same user showing variation with illiumination b)completely
saturated sample
Palm line variation with change in view
Parameter Sets
Base Image
FTABase
Image + 2
FTABase
Image + 6
FTABase
image + 10
FTA
Score Fusion
12.75% 0 24.30 0 19.25% 0 36.19% 0
T0 12.75% 0 13.99% 0 4.70% 0 5.79% 0
T1 7.64% 6.57% 8.07% 8.09% 5.35% 7.58% 4.5% 8.09%
T2 14.36% 9.44% 5.82% 13.15% 4.64% 10.9% 3.62% 11.8%
T3 3.69% 11.8% 3.40% 16.86% 1.9% 13.99% 7.75% 14.5%
•Dataset: 100 users, 6 videos each 640x480 logitech web camera used
• Combining 11 frames in the feature domain (using Gabor filter response as features) takes 1.4 secs on MATLAB•Parameters used for the Gabor filter experimentally determined to be the follwing: Window size = 27, var = 6.4, freq = 0.08• 3528 genuine comparisons and 1,75,065 imposter comparisons are made•Thresholds used on mean and variance of Gabor response of images to reject “bad” samples before running the algorithm
• First row shows results of score fusion as more frames are added• Rows 2-5 compare results while increasing quality control (threshold of expected minimum texture) for each of the combinations formed by adding 0, 2, 6 and 10 images respectively
Linear axis and Semilog axis ROC curve showing improvement in accuracies using multiple frames
ROC curve on semilog axis after rejecting
samples using threshold T3 as
shown in the Table.
It can be seen that the unusual drop in ROC is
lost, indicating that certain washed out
samples in data caused the problem
Frame Combination & Matching
• Frame combination is preferred over super resolution due to high computational complexity of the latter and a lack of robust landmark points in the low textured palm images • Average rule was experimentally found to perform better
• Matching: matching is done using Gabor filter responses.
Correcting for in-plane rotations and palm extraction
Frame 1 Frame 2 Frame 3
Technique used Base Image Base Image + 2 Base Image + 6 Base image + 10
Max. rule 18.24% 16.05% 19.34% 26.86%
Second Max. rule 18.24% 15.03% 15.48% 19.39%
Average value 18.24% 10.94% 9.55% 9.50%