دانلود رایگان ترجمه مقاله تشخیص حالت چهره غیر حساس از طریق بازسازی چهره سه بعدی (نشریه IEEE 2014)

تشخیص حالت چهره غیر حساس از طریق بازسازی چهره 3D با استفاده از فیلتر گابور از یک تصویر دو بعدی

Expression-Invariant Face Recognition via 3D Face Reconstruction Using Gabor Filter Bank from a 2D Single Image

Abstract

In this paper, a novel method for expression-insensitive face recognition is proposed from only a 2D single image in a gallery including any facial expressions. A 3D Generic Elastic Model (3D GEM) is used to reconstruct a 3D model of each human face in the present database using only a single 2D frontal image with/without facial expressions. Then, the rigid parts of the face are extracted from both the texture and reconstructed depth based on 2D facial land-marks. Afterwards, the Gabor filter bank was applied to the extracted rigid-part of the face to extract the feature vectors from both texture and reconstructed depth images. Finally, by combining 2D and 3D feature vectors, the final feature vectors are generated and classified by the Support Vector Machine (SVM). Favorable outcomes were acquired to handle expression changes on the available image database based on the proposed method compared to several state-of-the-arts in expression-insensitive face recognition.

1 Introduction

Expression-insensitive face recognition is one of the most difficult and challenging tasks in computer vision because of the changes in expression of human faces. However, few are focused on how to robustly recognize faces with expressions under the restriction of one 2D single training sample for each class. Available expression-insensitive face recognition methods can be mostly categorized into two separate types: 1) 2D-based techniques which use the 2D image to handle the expression in face recognition [1, 2] and 2) 3D-based methods which employ a 3D model (including depth and texture images) to handle expression in face recognition faces [3, 4, 5].

2D-based methods which are used from only a single image in the training set can be roughly divided into two main categories: model-based and optical flow-based. The basic idea in the model-based category is to warp images to similar global face transforms as the ones used for training. The concept of separately modeling texture and transform information has been applied in the Active Shape Model (ASM) and Active Appearance Model (AAM) [6, 7], and etc. Face transform is defined via a set of feature points in ASM, while face texture can be warped to the mean shape in AAM. Ramachandran et al. [8] presented preprocessing steps to convert a smiling face to a neutral face. Li et al. [9] applied a face mask for face transform normalization, and then calculated the Eigenspaces for transform and texture separately. However, this approach but not all images can be well warped to a neutral image because of the lack of texture in certain regions, like the openness of the mouth.