Using pattern recognition to glean meaning from masses of data is becoming faster and more accurate thanks to sophisticated algorithms and powerful, but economical, processors
Alan Joch, Senior Editor
An infant's first intellectual accomplishment is to recognize a parent's face from among the numerous people that walk into his or her life. Although the business world would benefit if computers could routinely perform the same skills, real-time pattern recognition with computers has traditionally been restricted to military applications and expensive supercomputers and mainframes.
Yet the value to the civilian world is obvious. Pattern recognition can help you classify and find meaning in masses of data, be it numerical, textual, audio, or video. The analysis techniques can al
so help you find matches between a target piece of data (e.g., a frame of video) and a database of millions of video images.
It's the underlying technology that makes today's pen systems recognize (or not) the written word. When you tell a computer to open a file in a pioneering speech-recognition system, it does so by matching your spoken words with a stored vocabulary of sounds. The quality-control systems that scan mass-production assembly lines for defective products find rejects thanks to pattern recognition. In time, cameras mounted on an ATM (automatic teller machine) may do more than just record your visit: A recognition system will match your face with a stored digital image to give you access to your bank account.
The following stories present three threads in pattern-recognition development. They illustrate how systems are becoming faster and more accurate.
Facial Recognition
CFR (computerized facial recognition) has been possible before today's generatio
n of systems, but the large computational tasks often took hours to complete even on the fastest hardware. Horsepower aside, a different facial expression, a new hairstyle, or differences in lighting often confused the algorithms written to match a ``live'' face with a reference image held in a computer's database.
In ``Face Value,'' Edmund X. DeJesus explores a new CFR system being deployed by the Commonwealth of Massachusetts. Built around an Alpha server and technology developed at MIT, the CFR system will hold the digitized faces of 4.2 million registered drivers. Within about 1 second, the state will be able to match a face with a digitized image. Unlike previous CFR systems, the Massachusetts implementation will be able to ``look'' past hairstyles and eyeglasses to make matches even when the digital facial images and the ``target'' image aren't exactly alike.
Key to the system's success will be its ability to select and store only the essential details that distinguish one person's face f
rom another. This will be important for making accurate matches and for keeping the storage requirements down to manageable levels. Filtering out all but the essential facial features, called eigenfaces, is also key to the system's fast response time.
So far, the program has proven to be quite accurate. In one test using a database of 7562 facial images, the program achieved a recognition rate of 95 percent. The immediate benefits for Massachusetts will be a crackdown on fraud by those who use duplicate licenses as false IDs. In addition, the facial database will streamline the process for drivers who need to replace a lost license. However, in the future, the same CFR system could create and search for stored digital images in multimedia databases. Soon, CFR-savvy computers may be smart enough to recognize their owners and automatically log onto a network, with all the proper security and access privileges, using facial verification rather than passwords.
Enabling Hardware
Real-time pattern recognition has been the domain of supercomputers and mainframes because each sample usually requires billions of recognition operations. Expensive hardware--and the custom programming that went along with it--slowed the growth of pattern recognition for civilian applications. However, business-class CPUs are now handling recognition tasks with the help of DSPs (digital signal processors) and neural-network processors.
In ``Eyes, Ears, & Brains on a Chip,'' Mark Clarkson talks to companies that are developing pattern-recognition applications around these hardware components. In one case, a company that developed a fingerprint-identification system replaced 28 circuit boards and four microprocessors with a single add-in board that holds twin DSPs. The cost for the two DSPs was about $800.
Similarly, a neural-network accelerator chip packaged within a development system, costing a total of $10,000, helped another company ship an OCR system that now reads 1000 characters per sec
ond, up from 15 cps in the previous version of the system. In the future, these processors can provide the scalable architectures and ability to work in multiple-chip implementations to meet future processing demands.
Patterns in Statistics
If tomorrow's pen-based computers become more accurate at recognizing handwriting, SPR (statistical pattern recognition) techniques will probably play a pivotal role in this increased accuracy. Handwriting recognition is one of a number of applications that depend on accurately classifying data, and classification is SPR's forte.
In ``Mining Statistics,'' John L. Cuadrado describes the underlying principles of SPR to explain how these techniques can efficiently tackle data-classification problems. SPR will evolve with theory-based classifiers to help doctors diagnose disease and help engineers avoid failure points in physical structures.
Piecing Together Puzzles
