| Summary: | 博士 === 國立臺灣科技大學 === 資訊工程系 === 104 === Various personal authentication systems have been extensively used in numerous civilian
applications because of the continual growth in the demands on security systems in
recent years. Biometrics has received considerable attention and has been extensively
used for identifying individuals in personal authentication systems. This thesis presents
a novel finger-vein recognition system based on the enhanced finger-vein images. To
achieve this, the system first identifies regions-of-interest from the captured images, and
then determines their skeleton topologies, which are used to analyze the similarities and
differences between finger-vein patterns. The system exhibited encouraging experimental
results in differentiating individuals, but failed in classifying some extreme cases of ambiguous
features. Consequently, an additional image quality assessment stage is borrowed
to enhance the recognition accuracy. As demonstrated in the experimental results, the
proposed extended strategy substantially improves upon the skeleton topology matchingonly
approach. The performance of the proposed method outperforms the existing systems
with the published databases and our own databases. The twofold examined finger-vein
recognition system exhibits great potential as a competitive biometric, and thus the practical
applications of which are vast. However, the captured finger-vein images are blurred
and low-contrast by the implemented low-cost near-infrared imaging device. Therefore,
the captured finger-vein image needs to be enhanced by the proposed contrast enhancement
method. Consequently, this thesis presents a novel local histogram equalization by
combining the transformation functions of the non-overlapped sub-images based on the
gradient information for edge preservation and better visualization. To ameliorate the
problems of the over- and under-enhancement produced by conventional local histogram
equalization, the bilateral Bezier curve-based histogram modification strategy is first employed
to modify the significant and insufficient changes of each cumulative distribution
in each sub-image. Yet, the gradient information has not been considered, and the cumulative
distribution of some enhanced sub-images are still significant or insufficient because
of the over- and under-enhancement, respectively. Therefore, the key insight of the proposed
method is that the transformation functions of the partitioned sub-images will be
weighed and combined based on the proportion of gradients to preserve the image texture.
In addition, the input image is separated into the non-overlapped sub-images for reducing
the time complexity. Based on the eight representative test images and mean opinion
score, the experimental results demonstrate that the proposed method is quite competitive
with four state-of-the-art histogram equalization methods in the literature. Furthermore,
according to the subjective evaluation, it is observed that the proposed method can also
apply to the practical applications and achieve good visual quality.
|
|---|
