通过低等级支持的极限学习机进行鲁棒高效的人脸识别资源-CSDN文库

156 浏览量 2021-04-12 23:16:35 上传评论收藏 2.56MB PDF 举报

人脸识别技术近年来在多种现实世界应用中取得了巨大进展，如身份验证、刑事侦查等。深度学习为视觉识别任务提供了一种端到端的范式，并且在性能上取得了良好的效果。然而，设计和训练复杂的网络架构耗时且劳动强度大。此外，在复杂的场景下，如光照变化、噪声干扰或遮挡等因素会降低识别算法的性能。为了解决这些问题，研究人员提出了一种名为低等级支持的极限学习机（Low-rank Supported Extreme Learning Machine, LSELM）的人脸识别算法，该算法在复杂场景下提高了识别性能，并且具有高效性。该算法分为三个层次：在第一层中，给定的样本被聚类到某些训练子空间中进行预聚类；在第二层，使用这个子空间，通过低等级分解恢复出一种对伪装、噪声、表情变化或光照变化具有抗干扰性的鲁棒特征；这些低等级判别特征被编码以支持训练一个前馈神经网络，也就是LSELM。通过实验结果表明，所提出的基于LSELM的人脸识别方法在某些广泛使用的人脸数据集（例如AR、Extend Yale-B、CMU PIE和LFW数据集）上与一些基于深度学习的算法在识别性能上不相上下，但其时间复杂度更低。具体而言，该算法在复杂场景下对伪装、噪声、表情和光照变化具有很强的鲁棒性。 LSELM算法相较于传统深度学习方法，主要的创新点和优势包括： 1. 高效率的训练过程：由于网络结构相对简单，LSELM的训练过程比深度学习算法更加高效。这意味着在相同的计算资源下，LSELM可以在较短的时间内完成训练。 2. 低等级特征的利用：LSELM算法通过低等级分解来获取鲁棒特征，即对各种环境干扰具有不敏感的特征。这使得算法即使在复杂场景下也能保持较好的识别性能。 3. 聚类与识别的结合：算法通过将样本进行预聚类，进而分层提取特征，并将这些特征用于训练前馈神经网络。这一过程结合了聚类技术和深度学习的优势。 4. 应用的广泛性：LSELM算法不仅在单一数据集上展现出良好的效果，而且在多种不同类型和复杂度的数据集上都有稳定的表现。 5. 对抗性攻击的抵抗能力：由于算法能够提取出不受伪装、噪声和光照变化影响的鲁棒特征，因此对于一些常见的对抗性攻击也具有一定的抵抗能力。 LSELM算法的提出为高效、鲁棒的人脸识别提供了一种新的视角和解决方案。它不仅适用于安全验证等传统领域，也为人工智能在更广泛的场景下提供了技术基础。随着算法的进一步优化和实际应用的检验，未来LSELM有望在人脸识别领域发挥更大的作用。同时，该研究也提示研究人员可以通过优化网络结构、改进特征提取方法，以及结合其他机器学习技术，来提高现有算法的鲁棒性和效率。

资源推荐

资源详情

资源评论

Robust and efficient face recognition via low-rank

supported extreme learning machine

Tao Lu

1,2

& Yingjie Guan

& Yanduo Zhang

Shenming Qu

& Zixiang Xiong

Received: 18 April 2017 / Revised: 23 November 2017 / Accepted: 29 November 2017 /

Published online: 23 December 2017

Springer Science+Business Media, LLC, part of Springer Nature 2017

Multimed Tools Appl (2018) 77:11219–11240

https://doi.org/10.1007/s11042-017-5475-2

* Tao Lu

lutxyl@gmail.com

* Zixiang Xiong

zx@ece.tamu.edu

School of Computer Science and Engineering, Wuhan Institute of Technology, Wuhan 430073, China

Department of Electrical and Computer Engineering, Texas A&M University, College Station, TX

77843, USA

School of Software, Henan University, Kaifeng 475001, China

Abstract Recently, face recognition algorithms have made great progress in various real-

world applications, e.g., authentication and criminal investigation. Deep-learning offers an

end-to-end paradigm for vision recognition tasks and achieves good performance. However,

designing and training the complex network architecture are time-consuming and labor-

intensive. Moreover, under complex scenarios, illumination change, noise or occlusion in

images degrade the performance of recognition algorithms. In order to ameliorate these issues,

we propose an efficient three-layered low-rank supported extreme learning machine (LSELM)

algorithm for face recognition which improves the recognition performance under complex

scenarios with high efficiency. In the first layer, a given probe sample is clustered into certain

training subspace as pre-clustering. In the second laye r, with this subspace, a low-rank

subspace of probe sample as robust feature which is insensitive to disguise, noise, v ariant

expression or illumination will be recovered by low-rank decomposition. Furthermore,

these low-rank discriminative features are coded to support training a forward neural

network termed LSELM. Experimental results indicate that the proposed approach is on

par with some deep-learning based face recognition algorithms on recognition perfor-

mance but with less time complexity over some popular face datasets e.g., A R, Extend

Yale-B, CMU PIE and LFW datasets.

Keywords Face recognition

Robust feature

Low-rank matrix recovery

Extreme learning

machine

Time complexity

1 Introduction

Face recognition (FR), which is widely used in information security [ 4]andpublic

safety, has been always a popular topic in computer vision and pattern recognition fields

[15, 36]. In recent years, it has gained a wide interest in both theoretical research and

industrial application domains. Although several face recognition products e.g., Deep

face [26], Face++, are released in last decade. It still seems a huge challenge in real

practice due to aging, occlusion, pose, illumination and expression variations and there is

a significant room for improvement [28, 32].

Generally speaking, there are two key steps of face recognitions, one is feature learning and

the other one is classifier. Benefiting from theory of computer vision, various hand-crafted

feature-extraction methods have been proposed and achieved good performances [20]. For

instance, Scale Invariant Feature Transform (SIFT) [39], Histogram of Oriented Gradient

(HOG) [35], Speed-Up Robust Features (SURF) [5], and Local Binary Patterns (LBP) [1]

feature-extraction methods, they are widely applied in traditional face recognitions. Consider-

ing the attribute of sample identity, subspace-learning based methods e.g., Eigenface [33],

Fisherface [6], and Locality Preserving Projections (LPP) [16] are developed to incorporate

label information from samples. Subspace-learning based algorithms are efficient and easy to

be applied to real applications. However, performance of abovementioned classical face

recognition algorithms drops when the input query images are disturbed by noise occlusions

or other outliers.

In order to promote the discriminative ability of classifier, by the merit of sparse represen-

tation theory, Wright et al. [37] proposed a robust facial recognition algorithm that achieved

satisfactory results even in noisy and complex conditions. This classifier, called as Bsparse

representation classifier (SRC)^, showed a strong discriminative ability by class-specific

reconstruction residual. Sparse representation randomly selects the dictionary atoms to repre-

sent the query input image, sometimes, results in unstable solution which degrades the

recognition performance. Some variants of sparse representation based approaches e.g., extend

SRC [10], discriminative SRC [41] were proposed to improve the stability of sparse repre-

sentation. Which enhances the discrimination of face recognition indeed, sparse or collabora-

tive representation? Some researchers argued that another kind of sparsity: locality often helps

face recognition. Collaborative representation based classifier (CRC) [40], used multi-category

atoms to collaboratively represent the query input image and distinguished its label informa-

tion from both reconstruction residual and the norm of coefficients. By the high efficiency of

ℓ

-norm, locality-constrained representation [24], probabilistic collaborative representation [7]

have been developed to furtherly promote the recognition performance.

Abovementioned representation-based classifiers and hand-crafted feature learning

schemes give impressive results. However, hand-crafted feature design schemes limit the

representation ability of massive data. Recently, deep learning based face recognition ap-

proaches as emerging machine learning paradigms become popular due to their powerful

representation ability. In order to close the gap between signal features and their semantics,

deep learning is used to automatically learn intrinsically features from large amounts of data

[30]. Different architectures of convolutional neural network (CNN) were investigated in [18]

11220 Multimed Tools Appl (2018) 77:11219–11240

to improve the performance of recognition, very deep networks were verified in [29]. Chan et

al. [9] proposed a simple yet powerful deep network using multi-layered principal component

analysis to enrich the discriminative ability during feature exacting. This algorithm was

considered as a baseline for deep-learning based face recognition on some famous face

databases, e.g., AR, extend Yale-B, CMU PIE and LFW datasets. These deep-learning based

approaches provide end-to-end solutions for big data application scenario. However, there are

two shortcomings for deep learning based face recognition algorithms: GPU cluster which is

essential for deep-learning optimization and large-scale date resources are always not available

for general researchers [2, 3, 17]. On the other hand, designing and optimizing these complex

networks are always time-consuming and labor-intensive. Above two shortcomings limit the

large scale extensive applications of deep-learning based approaches, especially in resource-

limited scenarios, e.g., mobile computing, autonomous robots.

In fact, developing efficient and accurate face recognition algorithms are still challenging

tasks. Especially, when the face images under complex conditions e.g., facial disguise, noise,

and expression or illumination changes, performances of feature extraction and classification

are both degraded. E. Candes et al. [8] theoretically proved that low rank minimization can be

used to rectify the defects in an observation matrix by decomposing it into two parts: low-rank

clear part and noise part. Thus this low-rank part of observation matrix would play more

important role in recognition against noise. Du et al. [11] proposed a low-rank constrained

sparse representation algorithm to enhance both feature representation and classification

abilities. However, it suffered from high computational complexity of low-rank and sparse

representation. Considering discriminations of traditional classifiers, e.g., K-Nearest Neighbor

(KNN) [34] and Support Vector Machine (SVM) [12], they are in efficient computing

manners, but just with limited accuracy. Some complex models including deep-learning based

approaches, e.g. FaceIDs [29, 30] and PCANet [9], achieve impressive performance, but they

are time-consuming to train their complex networks. Recently, a novel type of single hidden

layer feed forward network, called Bextreme learning machine (ELM)^ offers a fast training

paradigm for machine learning with strong generalization ability. ELM looks reasonable for

accelerating training process in face recognition scenario. But [19] had pointed out that even

though the ELM had advantages of low computational complexity and generalization ability,

its prediction accuracy was sensitive to the noise in the input data. Thus, when the training or

test data is noisy, the prediction performance of ELM drops dramatically.

To promote recognition performance with both high efficiency and accuracy [44], in this

paper, we propose a novel low-rank supported extreme learning machine termed LSELM,

aiming at extracting robust features, with fast learning manner and real application-oriented.

By the merit of low-rank recovery, the contaminated input images can be exactly separated

into clear inherent content and noise part. With those low-rank parts, the outliers would be

transferred into their inherent contents which bring discriminative ability of feature represen-

tation. Furthermore, ELM is utilized to accelerate training phase to get fast training results.

Consequently, by the low-rank content supported, the novel LSELM is robust to input noise

and time-efficient for face recognition. The main contributions of the proposed approach can

be summarized as follow: (1) We use low-rank recovery to decompose the input images into

inherent part and noise part, including variable illumination, disguise, and expression changes.

The low-rank part brings robustness of feature representation. (2) Low-rank supported extreme

learning machine, not only promote the robust representation ability, but also in an efficient

training manner. It ensures that the computational complexity of the proposed algorithm is

lower than other traditional approaches. (3) The proposed three-layered architecture is easy to

Multimed Tools Appl (2018) 77:11219–11240 11221

剩余21页未读，继续阅读

评论收藏

内容反馈

weixin_38617851

粉丝: 4
资源: 923

通过低等级支持的极限学习机进行鲁棒高效的人脸识别

高效的低等级支持的极限学习机，用于鲁棒的人脸识别

基于低秩约束的极限学习机高效人脸识别算法.pdf

基于局部线性嵌入极限学习机的人脸识别新方法.pdf

基于结构化遮挡编码和极限学习机的局部遮挡人脸识别.pdf

小波核极限学习机在人脸识别中的应用.pdf

基于加权KPCA和融合极限学习机的人脸识别.pdf

基于极限学习机与子空间追踪的人脸识别算法.pdf

基于投票极限学习机的人脸识别混合算法研究.pdf

基于多谐波极限学习机集成的人脸识别

脑网络启发算法：预训练的极限学习机

用于图像分类的极限学习机和自适应稀疏表示

lfw数据集 lfw人脸数据库.zip

神经网络案例分析

基于CORDIC的反正弦和反余弦计算的FPGA实现

使用3DCNN和卷积LSTM进行手势识别学习时空特征

BA无标度网络中的SIR模型

基于三次贝塞尔曲线的类汽车曲率连续路径平滑

基于机器学习的设备剩余寿命预测方法综述

基于FPGA的奇异值和特征值分解的快速实现。

基于BP神经网络的人口预测

磁悬浮系统自适应模糊PID控制器的设计

无人机协同目标的多无人机协同搜索方法

两轮平衡车的建模与控制研究

基于改进遗传算法的六自由度机器人时间最优轨迹规划

一种基于深度学习的机械臂抓取方法

最新资源