Facial Non-Recognition
At Least There's One Thing Humans
Are Better at Than Computers
I was sitting in a coffee shop the other day when a face I hadn't seen for more than a quarter-century crossed my vision. Given I was being harangued by my editor at the time, I couldn't give much of my attention to placing the face, but my subconscious did it for me: first name, school and class, plus some less relevant data such as the person's past ownership of a John Travolta-style disco suit. This is an extraordinary feat that our brain does constantly and in a nanosecond, and trying to make computers do it is a Holy Grail of the techie world.
Facial recognition is the ability to locate a human face, or faces, in a picture (whether it's a video or photograph) and then match the features to those stored in a database. That's pretty much what we do as humans; computers try to do the same thing, looking at, say, the relative distances between nose and mouth, or between the eyes.
The problem is, computers tend to get it wrong a lot. They are easily confused by poor lighting, a new mustache, or someone not standing properly in front of the camera. We humans are as good at recognizing someone in the real three-dimensional world -- from an angle, or moving -- as we are in a 2D picture. This is why I'm sitting in the cramped offices of Singapore's XID Technologies Pte., being shown a convincing video of Demi Moore's face doing pretty much whatever we want it to do. Her face revolves, showing a petite nose, sunken eyes and appealing mouth, with shadows passing over them. This convincing three-dimensional image isn't based on a video or lots of shots from different angles, but on one photo of Ms. Moore, culled from the Internet. "Usually you need to have many, many pictures of the individual" to make a believable 3D image of their face, says Roberto Mariani, XID's chief technology officer. "We used one."
XID software is now being used at a dormitory in Singapore, replacing fingerprint sensors with cameras that can quickly authenticate the complex's 6,000 migrant worker residents, even if they wear a hat, don't face the camera exactly straight on, or move when the camera is scanning them. In the first few weeks it did exactly what it was supposed to do, identifying 36 unauthorized men trying to sneak into the dormitory.
This is a sign that the software behind facial recognition is overcoming some of the issues that have kept it out of the hands of ordinary users. Another problem has been the computers themselves. Until recently, processors were simply not powerful enough to handle this kind of crunching. Now they are, meaning Web cameras can be used to authenticate users on laptops such as the SecureBook Fx9, which the Hong Kong-based RC Group launched in Asia recently.
Expect to see this kind of thing grow. But not necessarily on laptops. A more natural fit is the cellphone, which often already has a camera built in. Operators in Japan hope that facial authentication will boost cellphone-based commerce by overcoming customers' wariness about tapping their credit card details into their phone. "This is a way to drop dead any discussion of any security," says Lawrence Cosh-Ishii, co-founder of Wireless Watch Japan, a business intelligence Web site.
Facial recognition is likely to work the other way, too. As we take more photos, our computers can help us make sense of them. That's why Google recently bought a company called Neven Vision, which builds software to scan images for faces and then match them. Riya Inc. (
www.riya.com) already lets you upload your photos free and, after training the software with a few initial snaps, sorts them according to person. It works pretty well.
In the end, this sort of facial recognition may be more valuable than using it for authentication. More-sophisticated biometrics, such as those based on skin texture, may end up better for authentication, because they'll be more accurate and less vulnerable to fraud. But for identifying a face in a picture, facial recognition seems the way to go, whether it's organizing thousands of snaps or easing the night watchman's task of scanning faces and matching them to known undesirables. It'll never quite match us; this is one thing we're pretty good at.
人脸识别新境界
有一天我正坐在咖啡馆里，突然一张我有二十多年没见的面孔从我面前一闪而过。当时我正在听我的编辑侃侃而谈，没有办法仔细琢磨这张脸到底是谁，不过我的潜意识却悄悄地告诉了我答案，这人的姓氏、学校和班级、以及其他一些无关紧要的信息，比如他以前好像有一身像约翰?特拉沃尔塔(John Travolta)那样的迪斯科舞装束。这也是人类大脑一项了不起的功能，常常在瞬间就能完成上述的任务，但是要让电脑完成这样的工作对于科技业者来说就像寻找圣杯一样艰难。
人脸识别就是确定图像（录像或是照片）中人脸的形状特征，然后将这些特征与数据库中的资料进行比对。这与人类的行为颇为相似，电脑同样也要努力观察人的鼻嘴、或是双眼之间的相对距离。
问题是电脑经常容易犯错。光线不好、新留了胡子、或者在相机前站立的姿势不对等等都会让电脑不知所措。人类识别现实中三维立体人物形像的能力与识别照片中二维人物形像的能力一样强。为此，我坐到了新加坡XID Technologies Pte.局促的办公室里，看着戴米?摩尔(Demi Moore)逼真的视频头像按照我们的要求做出各种变换。随着她脸部位置的旋转，我们能看到她精巧的鼻子、凹陷的双眼和诱人的嘴唇。这副生动的三维图片并不是根据录像资料和不同角度的照片制作出来的，而是仅用了一张从互联网上找来的戴米?摩尔的照片。“为了制作逼真的人物脸部三维图像，往往需要获得这个人大量的照片，”XID的首席技术长(Roberto Mariani)说，“而我们只用了一张。”
XID的软件如今被用在了新加坡的一幢宿舍楼里，取代了原来的指纹识别系统。通过摄像头，这套软件能迅速识别出住在楼内的6,000名外籍劳工的身份，即便他们戴着帽子、或是没有正对着相机、或是处于走动的状态都没问题。在最初的几周，软件工作正常，认出了36个试图溜进楼里的外部人员。
这表明，人脸识别软件正在克服导致它远离普通用户的一些问题缺陷。而还有一些问题缺陷则和电脑自身有关。不久以前电脑芯片还无力处理这样的任务，但它们现在可以了，这意味着配置网络摄像头的笔记本电脑可以用来进行人脸识别，香港宏霸数码集团(RC Group)近期在亚洲推出的SecureBook Fx9即属此列。
类似的产品未来还将层出不穷，但不一定是和笔记本电脑相结合。由于现在的手机常常都配备了照相头，将人脸识别结合到手机便成了更自然的选择。日本的运营商希望，人脸识别技术能够推动手机增值业务的发展，让消费者无需再为在手机上输入信用卡信息而担心。“这能让所有安全方面的担心都烟消云散，”商业信息网站Wireless Watch Japan的联合创始人Lawrence Cosh-Ishii说。
人脸识别技术还能有其他用途。当我们拍摄了大量照片后，电脑可以帮助我们更好地管理它们。这也就是谷歌(Google)最近收购了一家名为Neven Vision的公司的原因。这家公司开发出一种扫描、匹配人脸图像的软件。Riya Inc.（
www.riya.com）已经可以让用户免费上传照片，在经过最初几张照片的识别训练后，软件就能根据不同的人对照片进行分类了，效果很不错。
不过，这种人脸识别技术在其他方面的使用价值可能要大于它的鉴别功用。基于皮肤文理等生物特征的识别技术可能更适合于完成鉴别工作，因为它们更精确、更不易出错。但是对于识别照片上的人脸，不论最终目的是用来整理数千张照片、还是为夜间值班员减轻工作负担，人脸识别技术似乎仍有着自己的优势。但不论怎样，人脸识别技术还是远远赶不上人脑本身，这方面人类还是很出色的。
Jeremy Wagstaff
