Weighted Gradient Feature Extraction Based on Multiscale Sub-Blocks for 3D Facial Recognition in Bimodal Images

Yingchun Guo,Yi Liu,Ruoyu Wei

doi:10.3390/info9030048

Abstract

In this paper, we propose a bimodal 3D facial recognition method aimed at increasing the recognition rate and reducing the effect of illumination, pose, expression, ages, and occlusion on facial recognition. There are two features extracted from the multiscale sub-blocks in both the 3D mode depth map and 2D mode intensity map, which are the local gradient pattern (LGP) feature and the weighted histogram of gradient orientation (WHGO) feature. LGP and WHGO features are cascaded to form the 3D facial feature vector LGP-WHGO, and are further trained and identified by the support vector machine (SVM). Experiments on the CASIA database, FRGC v2.0 database, and Bosphorus database show that, the proposed method can efficiently extract the structure information and texture information of the facial image, and have a robustness to illumination, expression, occlusion and pose.

Highlights

With the rapid development of technology and societal progress, efficient authentication is needed in many fields, e.g., surveillance, human–computer interaction, and biometric identification
Thakare et al [2] used the principal component analysis (PCA) components of the normalized depth image and moment invariants on mesh images to implement an automatic 3D facial recognition system based on the fuzzy neural network (FNN)
The FRGC v2.0 dataset consists of 4950 3D scans (557 subjects) along with their texture images (2D), in the presence of significant variations in facial expressions, illumination conditions, age, hairstyle, and limited pose variations

Summary

Introduction

With the rapid development of technology and societal progress, efficient authentication is needed in many fields, e.g., surveillance, human–computer interaction, and biometric identification. The improvement of 3D data acquisition on devices and computer processing capabilities make a rapid development of the 3D facial recognition technology. The existing 3D facial recognition techniques can be roughly classified into three categories: globally based, locally based, and multimodal hybrid methods. Global feature-based methods often extract statistical features from depth images [1]. Thakare et al [2] used the principal component analysis (PCA) components of the normalized depth image and moment invariants on mesh images to implement an automatic 3D facial recognition system based on the fuzzy neural network (FNN)

Methods

Results

Conclusion