公平人脸识别：数据平衡、增强与融合技术
需积分: 9 142 浏览量
更新于2024-07-14
收藏 256KB PDF 举报
"这篇论文《FairFaceRecognitionUsingDataBalancing,EnhancementandFusion》主要关注的是人脸识别领域的公平性问题，特别是在深度人脸识别中种族偏见的重要性。作者通过减少性别和肤色对识别结果的影响，提出了一种数据平衡增强和融合的方法来提高人脸识别的公平性。论文的作者包括Jun Yu, Xinlong Hao, Haonian Xie和Ye Yu，他们分别来自中国科学技术大学和合肥工业大学。该研究得到了中国国家自然科学基金和安徽省自然科学基金的支持。"
正文:
在当今的科技时代，人脸识别技术已经广泛应用于生活中的许多场景，如手机解锁、门禁系统以及视频监控等。然而，随着技术的发展，公平性问题逐渐浮出水面，尤其是当涉及到不同种族的人脸识别时。这篇研究指出，深度人脸识别系统可能存在种族偏见，这可能导致某些群体的识别准确率低于其他群体，从而引发社会公平性的问题。
为了缓解这个问题，作者Jun Yu等人提出了一种综合策略，包括数据平衡增强和融合两个关键步骤。数据平衡是指通过增加代表性不足的群体的数据量，使得训练数据集中各个类别的样本数量相对均衡，以减少模型对某一特定群体的偏见。这通常通过数据扩增技术实现，例如旋转、翻转或颜色调整等，以模拟不同的面部角度和光照条件。
增强环节则针对性别和肤色这两个可能引起偏见的因素。通过修改训练数据集中的这些属性，模型可以学习到更广泛的面部特征，从而降低性别和肤色对识别结果的影响。此外，融合策略则是在多个模型之间进行集成，每个模型可能对不同的特征有其独特敏感性，通过将它们的结果综合考虑，可以进一步提高整体识别的公平性和准确性。
论文还提到了中国国家自然科学基金项目(U1736123)和安徽省自然科学基金项目(1708085QF138)，这些资金支持了这项研究。作者们的工作展示了在人工智能领域，尤其是在人脸识别技术中，如何通过创新方法解决公平性挑战，这对于未来的技术发展和政策制定具有重要意义。
《FairFaceRecognitionUsingDataBalancing,EnhancementandFusion》是一篇深入探讨人脸识别公平性问题的研究，它提出的策略旨在减少因性别和肤色差异导致的识别误差，有助于推动更加公平、可靠的人脸识别技术的发展。这一领域的研究对于构建一个无歧视、包容的社会具有深远影响。
2 J. Yu et al.
African, Asian, and Indian. Similarly, gender is another aspect of face recognition
datasets imbalance, dataset mainly consists of male faces. To solve these prob-
lems, many efforts on face recognition aim to tackle the class imbalance problem
on training data. For example, in prior-DNN era, Zhang et al. [38] proposed a
cost-sensitive learning framework to reduce misclassification rate of face identi-
fication. To correct the skew of separating hyperplanes of SVM on imbalanced
data, Liu et al. [20] proposed Margin-Based Adaptive Fuzzy SVM that obtains
a lower generalization error bound. In the DNN era, face recognition models are
trained on large-scale face datasets with highly-imbalanced class distribution.
Range Loss [37] learns a robust face representation that makes the most use
of every training sample. To mitigate the impact of insufficient class samples,
center-based feature transfer learning [35] and large margin feature augmenta-
tion [33] are proposed to augment features of minority identities and equalize
class distribution. Besides, the FRVT 2019 [12] shows the demographic bias of
over 100 face recognition algorithms. To uncover deep learning bias, Alexander
et al. [3] developed an algorithm to mitigate the hidden biases within training
data. Wang et al. [32] proposed a domain adaptation network to reduce racial
bias in face recognition. They recently extended their work using reinforcement
learning to find optimal margins of additive angular margin based loss functions
for different races [31].
In this paper, we present a face recognition method to achieve fair face recog-
nition. Giving an image with a loosely cropped face roughly in the center, the
first thing we need to do is obtaining the face from the image. As a fundamental
step for various facial applications, like face alignment [28], parsing [7], recogni-
tion [34], and verification [9], face detection achieves great progress inspired by
deep convolutional neural network (CNN). Previous state-of-the-art face detec-
tors can be roughly divided into two categories. The first one is mainly based on
the Region Proposal Network (RPN) adopted in Faster RCNN [25] and employs
two stage detection schemes [29]. RPN is trained end-to-end and generates high
quality region proposals which are further refined by Fast R-CNN detector. The
other one is Single Shot Detector (SSD) [19] based one-stage methods, which
get rid of RPN, and directly predict the bounding boxes and confidence [8]. Re-
cently, one-stage face detection framework has attracted more attention due to
its higher inference efficiency and straightforward system deployment. We test
multiple face detectors, such as MTCNN [36], Retinaface [10], DSFD [17], and
propose an improved method based on DSFD.
After obtaining the face from image, as diversity between different attributes
is very large, a series of face preprocessing methods are used to reduce bias and
improve accuracy at the same time. For instance, we use a data re-sampling
method to balance the data distribution by under-sampling the majority class.
Train data enhancement and test time augmentation are used for obtaining im-
proved accuracy. Then train data would be used to train face recognition models,
trained mo dels are used to extract features of test data. Next, by calculating the
cosine similarity between two feature vectors, confidence scores of test data would
be generated, which indicate the degree two faces belong to the same person.
剩余14页未读，继续阅读
2024-10-18 上传
2024-10-18 上传
2024-10-18 上传
2024-10-18 上传
yico丶丶
- 粉丝: 2
- 资源: 11
我的内容管理
展开
最新资源
- 磁性吸附笔筒设计创新，行业文档精选
- Java Swing实现的俄罗斯方块游戏代码分享
- 骨折生长的二维与三维模型比较分析
- 水彩花卉与羽毛无缝背景矢量素材
- 设计一种高效的袋料分离装置
- 探索4.20图包.zip的奥秘
- RabbitMQ 3.7.x延时消息交换插件安装与操作指南
- 解决NLTK下载停用词失败的问题
- 多系统平台的并行处理技术研究
- Jekyll项目实战：网页设计作业的入门练习
- discord.js v13按钮分页包实现教程与应用
- SpringBoot与Uniapp结合开发短视频APP实战教程
- Tensorflow学习笔记深度解析：人工智能实践指南
- 无服务器部署管理器：防止错误部署AWS帐户
- 医疗图标矢量素材合集：扁平风格16图标(PNG/EPS/PSD)
- 人工智能基础课程汇报PPT模板下载
