Multi-Modal Ocular Recognition in presence of occlusion in Mobile Devices
Metadata[+] Show full item record
The existence eyeglasses in human faces cause real challenges for ocular, facial, and soft-based (such as eyebrows) biometric recognition due to glasses reﬂection, shadow, and frame occlusion. In this regard, two operations (eyeglasses detection and eyeglasses segmentation) have been proposed to mitigate the effect of occlusion using eyeglasses. Eyeglasses detection is an important initial step towards eyeglass segmentation. Three schemes of eye glasses detection have been proposed which are non-learning-based, learning-based, and deep learning-based schemes. The non-learning scheme of eyeglasses detection which consists of cascaded ﬁlters achieved an overall accuracy of 99.0% for VI SOB and 97.9% for FERET datasets. The learning-based scheme of eyeglass detection consisting of extracting Local Binary Pattern (LBP), Histogram of Gradients (HOG) and fusing them together, then applying classiﬁers (such as Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), and Linear Discriminant Analysis (LDA)), and fusing the output of these classiﬁers. The latter obtained a best overall accuracy of about 99.3% on FERET and 100% on VISOB dataset. Besides, the deep learning-based scheme of eye glasses detection showed a comparative study for eyeglasses frame detection using different Convolutional Neural Network (CNN) structures that are applied to Frame Bridge region and extended ocular region. The best CNN model obtained an overall accuracy of 99.96% for ROI consisting of Frame Bridge. Moreover, two schemes of eyeglasses segmentation have been introduced. The ﬁrst segmentation scheme was cascaded convolutional Neural Network (CNN). This scheme consists of cascaded CNN’s for eyeglasses detection, weight generation, and glasses segmentation, followed by mathematical and binarization operations. The scheme showed a 100% eyeglasses detection and 91% segmentation accuracy by our proposed approach. Also, the second segmentation scheme was the convolutional de-convolutional network. This CNN model has been implemented with main convolutional layers, de-convolutional layers, and one custom (lamda) layer. This scheme achieved better segmentation results of 97% segmentation accuracy over the cascaded approach. Furthermore, two soft biometric re-identiﬁcation schemes have been introduced with eyeglasses mitigation. The ﬁrst scheme was eyebrows-based user authentication consists of local, global, deep feature extraction with learning-based matching. The best result of 0.63% EER using score level fusion of handcraft descriptors (HOG, and GIST) with the deep VGG16 descriptor for eyebrow-based user authentication. The second scheme was eyeglass-based user authentication which consisting of eyeglasses segmentation, morphological cleanup, features extraction, and learning-based matching. The best result of 3.44% EER using score level fusion of handcraft descriptors (HOG, and GIST) with the deep VGG16 descriptor for eyeglasses-based user authentication. Also, an EER enhancement of 2.51% for indoor vs. outdoor (In: Out) light set tings was achieved for eyebrow-based authentication after eyeglasses segmentation and removal using Convolutional-Deconvolutional approach followed by in-painting.
Table of Contents
Introduction -- Background in machine learning and computer vision -- Eyeglasses detection and segmentation -- User authentication using soft-biometric -- Conclusion and future work -- Appendix
Ph.D. (Doctor of Philosophy)