异构人脸识别：SLBFLE方法的局部二进制特征学习与编码

107 浏览量更新于2024-07-15 收藏 1.05MB PDF 举报

本文主要探讨了一种名为"Simultaneous Local Binary Feature Learning and Encoding (SLBFLE)"的方法，针对的是异构和同构人脸识别问题。该研究由Jiwen Lu（IEEE高级会员）、Venice Erin Liong（IEEE学生会员）和Jie Zhou（IEEE高级会员）提出，他们旨在解决传统手工设计的面部特征描述符，如局部二值模式（LBP）和Gabor特征，通常需要大量先验知识的问题。 SLBFLE方法的独特之处在于它是一种无监督特征学习策略，能够在原始像素级别自动学习面部表示，无需预先设定的模板或规则。这种方法区别于诸如LBP、判别性面部描述符（DFD）和紧凑二进制面部描述符（CBFD）等两阶段特征提取过程，它将二进制代码的学习和代码书的构建结合在一起。这样做使得在处理不同身份的人脸图像时，能够通过一次性的特征学习和编码过程，有效地提取出具有区分度的信息。该方法的核心思想是通过耦合的同时局部二进制特征学习，捕获人脸图像中的关键模式和结构信息，这些信息对于识别不同个体至关重要。这种联合学习确保了特征的稳定性和鲁棒性，即使在光照变化、姿态角度和表情等因素的影响下，也能保持较高的识别性能。在实际应用中，SLBFLE可能包括预处理步骤，如图像归一化和局部区域选择，然后利用深度神经网络或者自编码器进行特征学习，生成二进制码。为了进一步提高准确性，可能还会结合某种形式的后处理，比如最近邻搜索或者支持向量机等分类器。相比于传统的手动特征工程，SLBFLE展示了更高的灵活性和适应性，因为它能自我适应并挖掘数据中的潜在模式，从而在异构人脸识别任务中展现出更好的性能。这项研究对于推进人脸识别技术的自动化和智能化具有重要意义，尤其是在跨设备和多模态识别场景中。

devices or environments because large appearance gap usu-

ally occurs. How to extract common properties and reduce

this gap is the key challenge in heterogenous face recognition.

Recently, a variety of feature learning methods [24], [25], [26]

have also been proposed for heterogenous face recognition.

For example, Jin et al. [26] learned representative features by

training a coupled ﬁlters which maximizes the inter-class var-

iations and minimize the intra-class variations. Saxena and

Verbeek [24] used a CNN model with a shared layer learned

from a soft-max criterion to obtain common features.

Yi et al. [25] extracted Gabor features from face landmarks

and performed shared representation learning to reduce the

modality gap. Differently, in this work, we learn speciﬁc and

common latent spaces to obtain similar information and

exploit speciﬁc complimentary information, respectively.

2.2 Binary Code Learning

A variety of binary code learning methods have been pro-

posed in recent years [27], [28], [29]. For example,

Weiss et al. [29] proposed a binary coding learning

approach for image search. Norouzi et al. [30] improved it

by using a triplet ranking loss optimization criterion. How-

ever, most existing binary code learning methods are devel-

oped for scalable similarity search [28]. While binary

features such as LBP and Haar-like features have been used

in face recognition, most of them are hand-crafted. There

have been some recent work which employs binary code

learning for face representation and recognition [31], [32],

[33]. For example, Zhang et al. [32] and Rastegari et al. [33]

learned binary codes based on variants of the ﬁsher crite-

rion. However, these binary codes are learned holistically

and not in feature level. More recently, Lu et al.[31] intro-

duced a compact binary feature descriptor (CBFD) which

learned binary face descriptors at the feature level. How-

ever, CBFD performed feature and codebook learning sepa-

rately, so that some useful information for codebook

learning may be compromised in the binarization stage.

3PROPOSED APPROACH

In this section, we ﬁrst review the LBP method and present

the proposed SLBFLE method. Then we show how to use

SLBFLE for face representation. Lastly, we present the pro-

posed C-SLBFLE method for heterogenous face recognition.

3.1 Review of LBP

LBP is an effective feature descriptor in face recognition [5].

For each pixel in face image, LBP ﬁrst computes the

difference between the central pixel and the neighboring

pixels and binarizes the difference with a ﬁxed threshold.

Second, these binary bins are encoded as a real value by

using a hand-crafted pattern coding strategy. Fig. 2 illus-

trates the basic idea of LBP, where two individual stages are

used for feature representation.

There are two shortcomings in LBP: 1) both the binar-

ization and feature encoding stages are hand-crafted,

which are not optimal f or local feature representation; 2)

a two-stage procedure is used in LBP, which is not effec-

tive enough because s ome useful info rmation for code-

book learning may be compromised in the binarizatio n

stage. To address this, we propose a SLBFLE method to

learn a discriminative mapping and a compact codebook

for f eatu re mapping and encoding jointly, so that more

data-adaptive information can be exploited in the learned

features. The following describes the details of the pro-

posed method.

3.2 SLBFLE

As aforementioned, our SLBFLE aims to jointly learn a fea-

ture mapping and a dictionary for feature mapping and

encoding. While our SLBFLE method is unsupervised, it

still has strong discriminative power because raw pixels are

extracted from face images of different identities which con-

tribute to learning a discriminative feature mapping. More-

over, the learned binary codes can well describe how pixel

values change over local patches and implicitly encode

important visual patterns such as edges and lines in face

images. Also, the learned dictionary can well encode the

learned binary codes so that some noisy information can be

well alleviated.

Let X ¼½x

; x

; ...; x

2R

dN

be a set of N face image

samples, where x

2 R

(1  n  N) is a pixel difference

vector extracted from an original face image. Fig. 3 illus-

trates how to extract a PDV for a given face patch. Com-

pared with the original raw pixel patch, PDV measures the

difference between the central pixel and the neighboring

pixels within a patch, so that it can better describe how pixel

values change spatially and implicitly encode important

visual patterns such as edges and lines in face images.

Assume there are K hash functions to be learned in SLBFLE,

which map and quantize each x

into a binary vector

¼½b

; ...; b

2f1; 1g

K1

, so that the binary codes

Fig. 2. The basic idea of the LBP method, where a two-stage procedure

is used for local feature extraction: feature mapping and feature encod-

ing. For the feature mapping stage, the difference between the central

pixel and the neighboring pixels are computed and binarized with a ﬁxed

threshold. For the feature encoding stage, the mapped binary codes are

encoded as a real value by using a hand-crafted pattern coding strategy.

Fig. 3. An illustration to show how to extract pixel difference vectors

(PDV) from the original face image. Given a face patch whose size is

ð2R þ 1Þð2R þ 1Þ, we ﬁrst compute the difference between the central

pixel and the neighboring pixels. Then, these differences are considered

as a PDV. In this ﬁgure, R is selected as 2, so that there are 24 neighbor-

ing pixels selected and the PDV is a 24-dimensional feature vector.

LU ET AL.: SIMULTANEOUS LOCAL BINARY FEATURE LEARNING AND ENCODING FOR HOMOGENEOUS AND HETEROGENEOUS FACE... 1981

剩余14页未读，继续阅读

weixin_38624746

粉丝: 3
资源: 946

异构人脸识别：SLBFLE方法的局部二进制特征学习与编码

局部二进制特征描述算法综述1

局部二进制模式

人脸图像特征提取matlab代码-multi_view-learning-summarize:我正在做一些关于多视图学习的研究，我想总结一下我

人脸识别实时性能挑战：快速响应解决方案

人脸识别门禁系统：OpenCV技术在其他领域的应用

PHP数据库同步与机器学习的协作：赋能机器学习模型的数据同步

MATLAB深度学习实战：构建神经网络模型，掌握深度学习技术，解锁人工智能的无限潜力

MATLAB 机器学习入门指南：解锁 AI 世界的大门

图像识别图像存储详解：深度解读图像存储技术在图像识别中的应用

Cell数组在图像处理中的应用：探索Cell数组在图像处理和计算机视觉中的强大功能

最新资源