Publication: Filter generation method in principle component analysis network for finger vein recognition
Loading...
Date
2024-09
Authors
Nurul Maisarah binti Kamaruddin
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The use of deep learning methods has gained significant popularity in machine learning, particularly in the field of biometrics. Various deep learning methods, such as deep neural networks (DNNs), convolutional neural networks (CNNs), AlexNet, and principal component analysis networks (PCANet), have been proposed to enhance biometric systems. Among these methods, PCANet has shown promising performance due to its simple design and effective use of basic data processing components like cascaded PCA, binary hashing, and block-wise histograms for image classification. However, PCANet's current filter generation approach does not account for specific characteristics of biometric images, such as the distinct vein patterns in finger vein recognition. This limitation means that the extraction of finger vein features may not be optimal, potentially affecting the overall accuracy of the biometric system. To address these shortcomings, this thesis proposes a new filter generation method for PCANet that takes into account important features of the image, such as vein lines in finger vein recognition. The proposed method utilizes two types of feature images: the original grayscale image and an edge feature image, to enhance the filter's sensitivity to vein patterns. Additionally, in the matching stage of the finger vein verification system, the choice of distance metric for comparing input biometric samples with stored templates plays a crucial role in the system's accuracy. Existing metrics may be limited in accuracy when applied to learned-based feature vectors. To overcome this challenge, this thesis also introduces a distance metric suitable for learned-based feature vectors to improve matching accuracy. The proposed methods were evaluated using three public finger vein databases: Finger Vein-Universiti Sains Malaysia (FV-USM), Shandong University Homologous Multimodal Traits (SDUMLA-HMT), and Tsinghua University Finger Vein and Finger Dorsal Texture (THU-FVT2). Experimental results indicate that the proposed filter generation method and distance metric significantly improve recognition rates compared to baseline methods. The classification accuracy of the proposed method shows improvements of 0.04%, 0.24%, and 0.33% over PCANet, and 0.95%, 1.57%, and 0.33% over CCANet for the FV-USM, SDUMLA-HMT, and THU-FVT2 databases, respectively. Additionally, the proposed distance metric achieved the best Equal Error Rate (EER) values, enhancing the overall verification performance. In conclusion, this research contributes an algorithm for filter generation and a novel distance metric to improve both identification and verification processes in finger vein biometric systems.