基于半核偏最小二乘法的面部图像超分辨率重建

研究论文

134 浏览量更新于2024-08-26 收藏 936KB PDF 举报

身份认证购VIP最低享 7 折!

30元优惠券

"这篇研究论文探讨了一种基于半核偏最小二乘法（Semi-Kernel Partial Least Squares, semi-KPLS）的面部图像超分辨率（Super-Resolution, SR）方法，用于实现面部幻觉。该方法分为两个步骤：首先，利用semi-KPLS算法构建非线性相关空间；然后，通过协同表示（Collaborative Representation, CR）来推断高分辨率（High-Resolution, HR）的面部图像。实验在FERET数据库上进行，表明所提出的算法相比现有的一些先进方法能产生更好的结果。面部特征的细节对于个体识别至关重要，因此提高面部图像的质量对于人脸识别技术有着重要的应用价值。" 本文的核心知识点包括： 1. **面部幻觉（Face Hallucination）**：这是一种技术，通过低分辨率（Low-Resolution, LR）的面部图像重建或生成高分辨率的面部图像，以增强面部特征的细节，通常用于人脸识别、监控和生物识别等领域。 2. **半核偏最小二乘法（Semi-Kernel Partial Least Squares, semi-KPLS）**：这是一种统计学习方法，它扩展了传统的偏最小二乘法，引入了核函数，使其能够处理非线性关系。在本研究中，semi-KPLS用于创建一个非线性的关联空间，帮助捕捉面部图像中的复杂结构和模式。 3. **协同表示（Collaborative Representation, CR）**：这是一种图像恢复或重建的技术，通过寻找数据集中的相似性来估计目标图像。在这个方法中，CR用于根据非线性相关空间中的信息推断高分辨率的面部图像。 4. **超分辨率（Super-Resolution, SR）**：是图像处理领域的一个关键概念，旨在提升图像的分辨率，增加细节信息，使图像更加清晰。在面部图像处理中，超分辨率可以改善人脸识别的准确性和可靠性。 5. **FERET数据库**：这是一个广泛使用的面部识别研究数据库，包含大量不同条件下的面部图像，用于测试和验证人脸识别算法的效果。 6. **实验比较**：论文通过在FERET数据库上的实验，展示了所提出的方法优于其他一些最先进的方法，这表明了semi-KPLS结合CR的面部幻觉方法的有效性和优越性。 7. **面部识别的重要性**：面部特征的细节是人脸识别的关键，尤其是在安全监控、生物识别系统和个人身份验证中。提高面部图像的质量有助于提高系统的识别性能和鲁棒性。通过以上知识点，我们可以理解这篇研究论文主要贡献了一种创新的面部图像超分辨率方法，结合了非线性建模和协同表示的优势，为面部识别技术带来了改进。

资源详情

资源推荐

FACE HALLUCINATION VIA SEMI-KERNEL PARTIAL LEAST SQUARES

Qiang Zhang, Fei Zhou, Fan Yang and Qingmin Liao,IEEE Member

Shenzhen Key Lab of Information Sci. &Tech., Shenzhen Engineering Lab of IS &DRM,

Department of Electronics Engineering / the Graduate School at Shenzhen, Tsinghua University,

China

zqtsinghua@126.com; flying.zhou@163.com; fyang@sz.tsinghua.edu.cn;

liaoqm@sz.tsinghua.edu.cn

Keywords-face super-resolution (SR); collaborative representation (CR); semi-kernel partial least

squares (semi-KPLS);

Abstract—In this paper, a patch-based super-resolution (SR) method is proposed to hallucinate

facial images. Two steps are involved in this method. In the first step, semi-kernel partial least

squares (semi-KPLS) algorithm is used to generate a nonlinear correlative space. In the second step,

we use collaborative representation (CR) to infer a high-resolution (HR) face. The experiments

conducted on FERET database demonstrate the proposed algorithm can generate better results, in

comparison with some state-of-the-art approaches.

Introduction

The details of facial features are crucial for identifying an individual from surveillance video, but

the resolution of a facial region from surveillance videos is usually very low. Therefore, in order to

obtain detailed facial features, it is necessary to infer a high-resolution (HR) face image from the

low-resolution (LR) one. This technique is called face hallucination or face super-resolution (SR).

According to the available inputs, existing SR methods can roughly be divided into two classes:

multi-frames SR [1-4], and single-frame SR [5-11]. The inspiration of multi-frames SR is to retrieve

high-frequency details from complementary multi-images. This type of method suffers from ill-

conditioned registration and is limited to small increase in image spatial resolution [15]. Instead,

single-frame SR prefers to infer the HR counterpart of a single image by obtaining extra

information from training samples and we mainly focus on this approach in this paper. In [14],

Freeman et al. propose an example-based method using a Markov Random Field (MRF) with belief

propagation, and HR images are generated by building a one-to-one relation between HR and LR

patches in training samples. In [18], based on the principle of locally linear embedding (LLE), HR

patches are constructed by a linear combination of several neighbors. However, most SR methods

are developed for the purpose of super-resolving general images.

For face SR, the utilization of the properties of face is conductive to generate the high-resolution

face images. Baker et al. [19] were the first to develop a hallucination method under a Bayesian

formulation and proposed the term “face hallucination”. In this method, it generates the high

frequency details from a parent structure with the assistance of training samples. This work is

further extended by Su et al. [23], where a steerable pyramid is used to provide larger search scope

for local best match approach. Eigen-transformation is introduced to hallucinating face in [5],

whose conclusion results in the example-based paradigm. That is hallucinated face can be

represented by a linear combination of training samples. In [24], canonical correlation analysis is

employed to maximize the correlation coefficients between LR and HR data. However, subspace

transformations may reduce the discrimination of hallucinated faces. Liu et al. [25] developed a

two-step approach integrating a global parametric model with Gaussian assumption and a local non-

parametric model based on Markov random field (MRF). Zhuang et al. [6] use locality preserving

projection and neighbor embedding to hallucinate a high-resolution face. More recently, the work in

[8] shows that patch-based method can achieve plausible results without the residue compensation.

In fact, the residue compensation is merely indispensable for the methods that include the step of

International Conference on Intelligent Systems Research and Mechatronics Engineering (ISRME 2015)

1143

下载后可阅读完整内容，剩余6页未读，立即下载

weixin_38626242

粉丝: 6
资源: 950

基于半核偏最小二乘法的面部图像超分辨率重建

最新资源