Abstract

In a multimodal biometric recognition system, the integration of multiple features derived from various biometric modalities seeks to overcome the several drawbacks found in a unimodal biometric system. In this paper, we have proposed a novel multimodal biometric recognition system based on a multi-level hybrid feature fusion mechanism to compact knowledge from multiple feature vectors. Several pre-trained networks with transfer learning, namely AlexNet, Inceptionv2, Densenet201, Resnet101, and Resnet-Inceptionv2, are employed to extract feature vectors to fuse with handcrafted feature vectors based on HOG feature descriptor. Canonical correlation analysis (CCA) and Discriminant Correlation Analysis (DCA) are utilized at a multi-level hybrid mechanism. To test the proposed framework, we used three biometric features: Ear, Face, and Gait. Numerical results have proved that our model outperformed other state of the art recent variants.

中文翻译：

---

基于多级混合特征融合的多生物识别系统

摘要

在多模态生物特征识别系统中，源自各种生物特征模态的多个特征的集成试图克服单模态生物特征系统中发现的几个缺点。在本文中，我们提出了一种基于多级混合特征融合机制的新型多模态生物特征识别系统，以压缩来自多个特征向量的知识。几个具有迁移学习的预训练网络，即 AlexNet、Inceptionv2、Densenet201、Resnet101 和 Resnet-Inceptionv2，被用来提取特征向量以与基于 HOG 特征描述符的手工特征向量融合。典型相关分析 (CCA) 和判别相关分析 (DCA) 用于多级混合机制。为了测试提议的框架，我们使用了三个生物特征：耳朵、面部和步态。