基于深度网络和多边形链距离的低质量视频人脸识别

需积分: 10 83 浏览量更新于2024-09-08 收藏 1.6MB PDF 举报

基于深度学习的低质量视频人脸识别技术在人脸识别领域中，低质量视频人脸识别一直是一个具有挑战性的问题。由于监控摄像头的普及，自动人脸识别技术对刑事调查的需求日益增加。但是，低质量视频的人脸识别仍然是一个亟待解决的问题。本文提出了一种基于深度学习的低质量视频人脸识别方法，旨在解决低质量视频的人脸识别问题。首先，我们提出了一个低分辨率的残差神经网络，作为人脸图像描述符。该网络通过质量调整的公共训练数据进行训练，数据通过数据增强策略生成，如运动模糊或添加压缩artifact。这种方法可以提高低质量视频的人脸识别率。其次，我们提出了一个基于多边形链的抗噪声人脸跟踪描述符，以进一步减少噪声效应。该方法使用多边形链来描述人脸特征，从而提高人脸识别的准确性。本文的贡献在于，我们提出了一个基于深度学习的低质量视频人脸识别方法，该方法可以解决低质量视频的人脸识别问题。我们的方法可以应用于刑事调查、身份验证、人脸识别等领域。知识点： 1. 低质量视频人脸识别的挑战：低质量视频人脸识别是一项具有挑战性的任务，因为视频质量的降低会导致人脸识别的准确性下降。 2. 基于深度学习的低质量视频人脸识别：我们提出了一个基于深度学习的低质量视频人脸识别方法，该方法可以解决低质量视频的人脸识别问题。 3. 残差神经网络：我们使用了一个低分辨率的残差神经网络作为人脸图像描述符，该网络可以学习到人脸特征。 4. 数据增强策略：我们使用数据增强策略来生成质量调整的公共训练数据，如运动模糊或添加压缩artifact。 5. 多边形链：我们提出了一个基于多边形链的抗噪声人脸跟踪描述符，以进一步减少噪声效应。 6. 应用场景：我们的方法可以应用于刑事调查、身份验证、人脸识别等领域。本文提出了一种基于深度学习的低质量视频人脸识别方法，该方法可以解决低质量视频的人脸识别问题，并且可以应用于多个领域。

Low-Quality Video Face Recognition with Deep

Networks and Polygonal Chain Distance

Christian Herrmann

∗†

, Dieter Willersinn

†

, J

urgen Beyerer

†∗

∗

Vision and Fusion Lab, Karlsruhe Institute of Technology KIT, Karlsruhe, Germany

†

Fraunhofer IOSB, Karlsruhe, Germany

{christian.herrmann|dieter.willersinn|juergen.beyerer}@iosb.fraunhofer.de

Abstract—Face recognition under surveillance circumstances

still poses a signiﬁcant problem due to low data quality.

Nevertheless, automatic analysis is highly desired for criminal

investigations due to the growing amount of security cameras

worldwide. We suggest a face recognition system addressing the

typical issues such as motion blur, noise or compression arti-

facts to improve low-quality recognition rates. A low-resolution

adapted residual neural net serves as face image descriptor. It

is trained by quality adjusted public training data generated by

data augmentation strategies such as motion blurring or adding

compression artifacts. To further reduce noise effects, a noise

resistant manifold-based face track descriptor using a polygonal

chain is proposed. This leads to a performance improvement

on in-the-wild surveillance data compared to conventional local

feature approaches or the state-of-the-art high-resolution VGG-

Face network.

I. INTRODUCTION

The increasing availability of security cameras raises the

demand for analysis of the vast amounts of video footage,

speciﬁcally, automatic analysis because manual inspection is

unfeasible. While older security cameras lack in resolution and

faces are often unrecognizable, with newer camera generations

the faces become clearer and distinguishable. However, the

data quality is usually still far from professional footage such

as TV or press photographs, where automatic face recognition

achieved impressive results recently, surpassing even human

performance in certain setups [1, 2]. Addressing the low-

quality surveillance domain is still a signiﬁcant challenge

for automatic face recognition approaches, caused by sev-

eral reasons which are misalignment, noise affection, lack of

effective features and dimensional mismatch between probe

and gallery according to [3]. Recently, effective alignment

methods for low-quality faces were proposed [4] which we

found to be sufﬁciently accurate. Consequently, in this paper,

we suggest an effective Convolutional Neural Network (CNN)-

based feature which proves to be more efﬁcient compared

to previous solutions and address noise affection by data

augmentation and a noise resistant track descriptor to utilize

the temporal information. Data augmentation is necessary

because large face datasets which are suitable for training a

CNN are no surveillance datasets and consequently involve

a domain gap. We suggest according augmentation strategies

such as adding motion blur or compression artifacts to close

this gap.

Found a solution

Config

21 55 12 86

11 6 16 16

Similarities:

1.2429 1.7796 1.4668

0.8777 0.7699 0.9891 0.9272

1.7920 1.8887 1.3830

0.93 0.99 0.77 0.88

1.25

1.79

1.78

1.89

1.47

1.38

Fig. 1: Qualitative results of the proposed method on low-

quality data. Line thickness and numbers denote face similarity

(inverse of descriptor distance) and line color same (blue) and

different (orange) identity.

In detail, the contributions of this paper are threefold:

First, the adaptation of the residual net architecture [5] to

the low-quality face recognition domain by adjusting layer

conﬁguration and setup. Second, a manifold based and noise

resistant strategy using a polygonal chain to aggregate facial

information across multiple frames of a face track. Third, a

systematic analysis of the target domain image quality effects

and their reproduction as data augmentation for high-quality

training data from a different domain.

II. RELATED WORK

Addressing low-quality video face recognition involves sev-

eral speciﬁc problems.

Face recognition with CNNs. Currently, CNNs serve only

for high-resolution face recognition where they signiﬁcantly

improved the performance compared to previously known

approaches, even surpassing human capabilities in certain

setups [1, 2, 6]. Because these networks are mainly based

on solutions for the ImageNet challenge [7], they adopt the

high resolutions of 224×224 pixels and above. Low-resolution

networks tend to loose performance as shown by [1], which

makes it necessary to address this issue.

Low-quality face recognition. One part of the approaches

addressing this task tries to mitigate the low data quality

by preprocessing steps. This includes super-resolution meth-

ods where low-quality images are upsampled to apply a

conventional high-resolution face-recognition strategy [8, 9]

which proved to be a solid strategy for comparing low-quality

下载后可阅读完整内容，剩余6页未读，立即下载

KONGPEILING

粉丝: 0

基于深度网络和多边形链距离的低质量视频人脸识别

Face Recognition简易实现：结合face_recognition和PyQt5

Nextcloud中的面部识别应用：FaceRecognition功能介绍

基于Face Recognition与Flask的人脸识别服务教程

【Advanced Level】Image Face Recognition in MATLAB: Using FaceNet for Image Face Recognition

Impact of Python Versions on OpenCV Image Recognition Accuracy: Experiments and Analysis, Enhancing ...

【Introduction】: Demystifying the Principles of Generative Adversarial Networks (GANs): Essential ...

: A Major Contest of GAN Architecture Performance: Who is the Pioneer of Deep Learning?

face_recognition 1.0.0版本安装指南

"深度人脸识别综述：最新进展与关键元素

cole_02_0507.pdf

最新资源