Wavelet Decomposition Feature Parallel Fusion by Quaternion Euclidean Product Distance Matching Score for Finger Texture Verification - Advances in Computational Science and Engineering - page 139

algorithms are proposed and the ideas are used in transform-domain of original
image [5].
(b) Fusion at feature level: This intermediate-level fusion computes feature
data from each sensor into a feature vector, for instance, each feature vector from
sensor can be concatenate into a vector. It gains a better recognition performance
in theory than other fusion levels due to their storage of original information
from data. But it usually raises problem of space incompatibleness, or curse of
dimensionality.
(c) Fusion at matching-score level: Matching scores are provided by system
indicate proximity of the feature vector with the template vector for personal
identification or authentication [6].
(d) Fusion at decision-making level: This high-level fusion makes a fusion
after matching of each trait. That is, it combines the classification result of
each feature vector, i.e. accept or reject [6]. As a result, the fusion has a high
robustness and low algorithm complexity.
Recently, it raises a heated research on feature level, i.e. Feature Fusion Level
(FLF) in the community of multimodal biometric recognition, because of (a)
richer information which can be acquired from raw data than other level fu-
sion and (b) good performance without an independence assumption rather
than decision-making level. And FLF classifies into four aspects based on fusion
method, namely serial fusion, parallel fusion, weight fusion, and kernel fusion.
Traditionally, serial fusion is a most common used fusion method, yet it has
two evident drawbacks: (1) the dimensionality of feature vector after concate-
nation increases dramatically so that recognition speed goes down sharply. (2)
as to Small Sample Size (SSS), e.g. face recognition, this method usually gives
rise to singularity of the within class scatter matrix with a high dimensional-
ity of concatenated feature matrix, which is an obstacle of feature extraction of
Fisher criterion. To solve such problems, [7,8] proposed parallel fusion with a
complex feature vector form that aims at obtaining a higher recognition effect
and a faster speed than serial fusion. According to the advantage above, we pro-
pose an approach of 2D wavelet decomposition feature parallel fusion in terms
of a characteristic of QEP for finger texture authentication. We propose a new
finger texture ROI as our extraction ROI with more information for wavelets
feature acquisition but at a low dimensionality. Then partial wavelet coecients
by level 2 wavelet decomposition are obtained as fusion features so as to avoid
redundant feature data like all the coecients used. After that, we use QEPD, a
distance between modulus square of template quaternion and modulus of tester
QEP as matching score for finger texture verification. For proving the effective-
ness of experiment of finger texture QEPD, a comparison with palmprint QEPD
is conducted. The final recognition result with 93.36% is higher than the same
means with 92.87% by palmprint. In addition, the dimensionality of original
image sample, composed quaternion matrix the system processes for finger tex-
ture, are lower than palmprint. Therefore, we can conclude such finger texture
is a better biometric than palmprint for QEPD. Also, it is a fast time for the