الفهرس 
MotivationsOnly 14 pages are availabe for public view
 |  | from | 151 |  |  |
| | | from | 151 | | |
Abstract
Biometrics such as face, fingerprint and iris are defined as the physiological characteristics of an individual. It has become a need now to get more affordable biometric systems that could be easily embedded in smart devices such as tablets and mobile phones. A human face is considered as one of the most effective biometric traits compared to other biometrics as it has high acceptability during acquisition. In the last few years, researchers who are interested in the field of face recognition (FR) proved that the utilization of convolutional neural networks (CNNs) gives more robust, representative, and detailed features that improve the overall system performance. Paying attention to enhancing the recognition accuracy should not make us overlook the protection of biometric data from hackers. Thanks to cancelable biometric techniques, biometric data protection can be provided with a slight degradation in the recognition accuracy. In this work, some cancelable CNN-based FR methods are proposed; the single-region FR method, the region-based FR method, the hybrid-features FR method, and the multi-biometric FR method. In the single-region FR method, a single CNN extracts deep-learned features from face images to obtain an appropriate and discriminative facial descriptor. Furthermore, an efficient CNN model is proposed. This proposed model exploits batch normalization, depth concatenation (DC), and a residual learning framework.
Abstract
II
The region-based FR method includes the detection of; face, eyes, nose and mouth regions from the original face images. Multiple CNNs are used to extract deep features (DFs) from each region. The extracted features can be combined using a fusion technique (FT). Various techniques can be used to perform fusion such as principal component analysis (PCA), discrete wavelet transform (DWT), multi-set discriminant correlation analysis (MDCA), and a fusion network. The hybrid-features FR method explores the merits of both DFs and hand-crafted features. A CNN is used to extract the DFs from the face images. Moreover, hand-crafted features can be extracted using traditional techniques such as speeded up robust features (SURF), scale invariant feature transform (SIFT), local binary pattern (LBP), oriented rotated BRIEF (ORB), and histograms of oriented gradients (HOG). In addition, to reduce the dimensions of hand-crafted features to be consistent with the dimensions of the deep-learned ones, a dimensionality reduction (DR) method can be applied using PCA technique or independent component analysis (ICA). Finally, the DFs and hand-crafted features are combined through a fusion process. The multi-biometric FR method consists of three phases. The first phase includes data collection from various biometric traits such as face, iris, fingerprint, palm print, and ear. The second phase incorporates the use of five CNNs to extract features from the different traits. Finally, a fusion network combines the extracted features to get an appropriate and sufficient facial descriptor.
Abstract
III
For all the proposed methods, to provide user’s privacy and increase the
system resistance to spoof attacks, a cancelable biometric technique is applied on
the final facial descriptor. Cancelability can be provided using different methods
such as bio-convolving and bloom filter. These methods use one-way function to
transform the biometric templates instead of storing the original ones. In
addition, the recognition accuracy is slightly degraded as the statistical
characteristics of features after transformation are approximately maintained.
The experimental results on various datasets prove that the proposed CNN
model achieves remarkable results compared to the other state-of-the-art CNNs,
the region-based FR method is superior to the other proposed methods, and the
bio-convolving method performs better than the bloom filter method in
providing cancelability.