To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

Author: intellectual interest in cloud based visual passenger identification AI, AI focused scene recognition service providers


According to estimates of the Institute for Prospective industries face recognition market, by 2022, China will face recognition market size of more than 6.6 billion yuan

To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

About Face a variety of fragmented reports, they are endless, but few articles can be systemic, from the application point of view to explain the depth of face recognition.

Intellectual interest as a focus on customer identification cloud scene recognition AI service providers, we will use technology and business model in two dimensions, and strive to give readers, especially considering practitioner AI + application, see the whole picture.

Manual carefully before speaking we face recognition technology.

1 base layer algorithm

We may hear intermittent extraordinary face detection, face recognition algorithms, but whether there are links between these algorithms, whether there are system? The answer is "some."

We can face recognition algorithm is divided into a base layer and application layer algorithms.
To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

Base layer algorithm, pre-processing corresponding to a human face. A human face, we must first go through face detection, feature key point processing, quality after filtration model, in order to make the application layer processing algorithm, and applied to the actual scene.
Pros and cons of the base layer algorithm, will largely affect the final recognition accuracy and effectiveness.

technology definition effect principle
Face Detection Insert a photo or a video stream out face detection, and output face matrix coordinates For intercepting face for subsequent alignment of face, a face search algorithms. Dichotomous model, the depth of learning by training samples whether a human face
Key features After detection of a human face, the face feature points marked, each feature point has attributes, the face position can be represented by 1 straightened face alignment: the actual scene, captured face generally not the positive direction, than human face, the face search and other processing needs <BR> 2 and then straightened: Mutual entertainment applications such as stickers after facial effects, the need to detect facial feature key, the targeted treatment of key parts The key points of the face are Mark the photo, by the depth of learning, classification model, so that the algorithm can detect feature points and feature points of the recognition properties.
Quality Model Angle photo of the human face, light, blur, etc. to assess, so that photos meet the requirements of recognition to the next step 1 适应不同业务中对照片的需求,比如有些场景,需要口罩能识别(医院),而有些场景则不能<BR>2 提升人脸比对、人脸搜索等后续人脸识别的准确率 回归模型,对每张照片标记模糊、光照、遮挡的分值,进行监督训练后,输入照片即可输出对应的质量分值。

To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

2 应用层算法

目前,人脸识别在身份认证领域与互娱领域应用最为广泛;在智能交互,数据分析处理等方向上,人脸识别也在进行着积极探索。

身份认证/安防的核心功能在于确认“你是谁”,互娱领域的核心在于“人脸特效处理”;两个领域,两条赛道,分别拥有各自不同的产业链。

身份认证犹如一位思维严谨的工程师,狠抓识别准确率,防***等指标,并结合应用落地场景,串联业务流程,也是当下AI结合产业互联网的典型。

互娱领域就像一位钻研人性的产品经理,打造各种人脸特效,美颜、贴纸等都不在话下,并结合平台用户偏好,使用针对性的人脸特效策略,引领甚至塑造人们的审美潮流。

目前,智趣云识客专注于第一条赛道上的核心技术,用心服务上下游企业。虽然这会使得我们的业务面没有那么全,但是却能让我们在特定赛道上占有一席之地,关键核心技术已达到行业领先水平。

To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

2.1 你是谁?无介质证明身份

日常生活中,原来我们都是需要通过介质(×××、工牌、驾驶证等)来证明身份,而以人脸识别为代表的生物识别,则无需介质。

身份认证/安防的核心技术在于活体检测、人脸比对、人脸搜索;主要用于:线上远程认证场景(金融开户、刷脸注册、刷脸登录等)、线下无人值守场景(智慧交通、人脸门禁、刷脸取款、刷脸支付等)。

  • 活体检测
    是身份认证的第一步,因为首先我要确认这个人是真人,而不是视频、照片、面具等欺诈盗用行为。

活体检测的技术上,目前也主要有两大类:对硬件依赖度比较低的,如动作活体,静默活体;对硬件有一定要求,需要和硬件适配的,比如双目活体、3D结构光活体等。虽然后者的成本比前者高,但是防***效果更好,而在线下场景中,天然的需要硬件,因而后者也成为线下场景的最好选择。

原理上,都是采集人脸照片,并将照片做上标记(真/假样本),并送到模型中训练从而得出算法。不同的活体检测,因为样本源不一样,比如红外摄像头采集的照片,带有灰度特征;3D结构光采集的照片带有深度信息,导致识别效果也不同。所以,活体检测的关键,除了算法、模型构造,还有一个就是图片样本本身所带有的信息量。

To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

  • 人脸比对
    是将两张人脸照片进行比对,得出相似度;第一张是现场采集的,第二张该如何得来?一般有两个来源:

1 另外一个能代表你身份的载体,比如×××、行驶证、驾驶证等证件照,这类场景用来做金融开户、人脸注册、网约车司机认证等场景,通过现场采集照比对你的证件照信息,确认你就是本人。

2 账号下已经绑定的人脸:一般需要先输入账号,获取对应人脸。这类场景的典型应用是取代原来的密码功能,比如刷脸登录、刷脸支付等。

To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

  • 人脸搜索
    是将采集到的人脸,和底库中的人脸全部进行比对,得出相似度最高的几张人脸底库照,并得出相似度,超过一定阈值,则可以认为是同一人。

人脸搜索,无需事先得到人脸照,只需要刷脸即可,在线下门禁等安防领域,线下刷脸支付等应用广泛。当然,不同的业务领域中,根据误识的后果,对人脸搜索的容错性也不一样;比如在工地人脸识别中的容错率,就要比在刷脸支付中的容错率要低。

需要说明的是,人脸搜索的准确率,是要结合人脸底库中人脸照片的数量来的,底库中人脸照片越多,识别准确率越低。这个和人一样,在2~3个人中,找出你曾经认识的人,比较容易;但是上百万个人,则长相相似的人也越多,辨识更困难。目前业界做的好的一般是百万级别的人脸库,识别准确率在95%以上

To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

2.2 从工具到社交,娱乐至上

互娱应用,也深深契合着行业发展。

起初随着智能手机兴起,人们的自拍分享需求渐渐旺盛,美颜滤镜,作为与手机硬件深度结合的产品,见证着人们变美的时代,此时,算法主要由第三方算法公司提供。

随着4G时代带来,短视频社交成为人们生活热点,美颜滤镜、贴纸也应用于各大互娱平台中,并成为不可分割的一部分;对于短视频内容生产者来说,甚至已成为核心竞争力。因此,诸如快手、抖音等平台,都以自研算法,并结合客户群画像,独自研发。

滤镜是图像美化中必不可少的步骤, 所谓滤镜,最初是指安装在相机镜头前过滤自然光的附加镜头,用来实现调色和添加效果。2008年,美图一炮而红,人们发现,原来滤镜还可以这么玩,自此,美颜滤镜开始了从工具到美学定义者的转变。

早起的传统算法,主要是先使用人脸特征关键点算法,勾画有效区域,然后在不同的区域进行亮度提升、去噪声等算法,实现美颜滤镜。

随着深度学习的兴起,研究人员们开始更关注结果,设计师将原图P成美化完成后的结果图,并用于训练。人们美颜后,究竟想变成什么样?研究重心也开始偏移。

Technical Level

贴纸,人脸融合,则是更高阶的玩法。核心还是人脸特征关键点,对于贴纸和人脸融合来说,关键点的数量越多越好,对齐的越准确。人脸融合,则是将两张人脸的关键点进行融合。

Technical Level

2.3 不断进取,跨越感知智能

人工智能承载了业界对于世界改造的期望,一定程度上说,属性识别、视线估计、gan等,从感知智能程度上往前更进了一步,但是因为技术不够成熟、商业应用领域狭窄等原因,至今未得到大规模商业应用。可以说,视觉AI想跨越到认知智能,AI与AI之间相互融合,依然还有很漫长的路要走

  • 属性识别
    识别年龄、性别,高兴、悲伤、愤怒等情绪,获取用户更多维的数据,丰富用户画像、个性化推荐、广告展示等等场景,听着很美好,对不对?毕竟在数据为王的时代,数据就是价值。但是,商用化还是存在技术硬伤,识别准确率也就70%左右。

近日,美国等5名专家,耗时两年,查阅1000多项研究,在论文《再论人类情感表达:从人类面部表情辨别情绪的方法论面临的挑战》(论文原名为:《Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements》)中表示:人类情绪的表达方式及其丰富复杂,很难靠简单的面部表情识别,人们生气时,在平均不到30%的时间里他们会皱眉,故皱眉不等于愤怒,皱眉知识“愤怒”的众多表达方式之一。同时,表情和语言、情境的相关关系也非常大。

To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

  • 视线估计
    和人脸特征关键点比较像,检测完人脸之后,再检测人眼以及眼球,并锁定眼球中心等关键点位置,根据坐标来锁定视线方向。主要应用于课堂上,评估学生注意力;AR VR等新型硬件交互,通过视线方向,自动切换视频中的位置等;广告投放,评估行人对广告的注意力;目前而言,市场体系还是比较小,未得到大规模应用

To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

  • GAN
    全称为生成对抗网络,初衷是生成不存在于真实世界中的数据,使得AI具有创造力或者想象力,也是目前AI领域一个比较热门的研究方向。

gan的核心网络分为生成器与判别器;生成器负责凭空捏造数据,判别器负责判断数据是否是真数据;两个核心网络相互博弈,直至动态平衡,让生成的数据无限逼真与真实数据。

To apply the perspective of a comprehensive analysis of face recognition | intellectual interest in cloud customer knowledge

如图,随机噪声就是随机生成的一些数,也就是gan生成图像的源头。

生成器根据一串随机数生成一个假图像,并用这些假图去欺骗判别器

而判别器通过真图和假图的数据(相当于天然的label),进行一个二分类神经网络训练,并判别输入的是真图还是假图,给出一个分值。

For example, FIG real face is a series of photos. Initially, the generator generates a photograph, is certainly a mess, but to judge scoring discrimination will tell generator, you really are not generated map (facial photo), so the generator according to the depth of learning, such as back-propagation, continue to modify their own pictures, and then, the resulting picture will be more close to real human face, until homeostasis.

gan attention, for many reasons, such as:
1 itself is unsupervised. At present, most commercial applications of artificial intelligence algorithms are able to supervise algorithm, the so-called supervision algorithm, is the need for vast amounts of sample, and manual tagging, people learn to tell the depth of the network is correct, and dissemination of training, so the industry has "had how much labor, there are that many smart "ridicule.
2 AI have let the imagination, such as the fuzzy graph becomes clear (to the rain, to fog, to shake, to the mosaic), the brain can make the plot
a lot of paper in the study of the development prospects of gan.

Written in the last 3

Any technology, also followed the technical development -> mature technology -> law of development of commercial landing

Pond technology innovation, but also from business technology pool, explore appropriate technology to transform the world;

Face recognition as a complex technology, both have now, and also continue to develop in the future. Though there are difficulties, but the outlook is exciting.

Guess you like

Origin blog.51cto.com/14420407/2428861