Computerized Face Recognition
A new method for computerized facial recognition is faster and better than ever. Researchers believe this revolutionary technology may make a big difference for security systems, since it can recognize faces in spite of disguises, state the authors of research published in the July 2009 issue of International Journal of Intelligent Systems Technologies and Applications.
Researcher Lin Huang, of Boca Raton's Florida Atlantic University explains that while each face has a defining feature, faces can still be very much alike, making the development of facial recognition technology a daunting task for security systems and for other spheres where the technology has applications. Many years have been spent trying to develop and refine facial recognition software, yet the technology is still not in common use for such purposes as identification at border crossings, gaining access to buildings, automated banking, and criminal investigations. The major issue throwing a spanner into the works is the fact that these systems need a great deal of computer power to do their magic.
The earliest facial recognition systems measured the distances between prominent facial features: eyes, nose, and mouth as displayed in photographs. During the1970's, the process became more automated through the use of templates and mapping. With the 1980's came the first completely automated facial recognition system which relied on a statistical approach.
At the end of the 1980's Brown University researchers came up with the "eigenface method," which was expanded upon by MIT scientists during the first part of the 1990's. Since this time, newer approaches have focused on a variety of models, for instance neural networks, dynamic link architectures (DLA), fisher linear discriminant (FLD), hidden Markov, and Gabor wavelets. Finally, a technique was developed that could create an image resembling a ghost and which could be subjected to a more in depth analysis so as to spot the major differences in various faces.
The last hurdle to overcome has been the necessity of powerful computers to run the software. To this end, Huang, and partners Hanqi Zhuang and Salvatore Morgera from the Department of Electrical Engineering, have found a way to apply a one-dimensional filter to the two-dimensional data generated by the usual analyses, such as, for instance, the Gabor method. This means they can cut the required amount of computer power without sacrificing speed or accuracy.
The three researchers tried out the algorithm using a typical database of 400 images for 40 subjects. Images appear in grey scale and are small at 92 x 112 pixels. Besides being faster, the technique can work with low resolution images, such as are produced by typical CCTV cameras. In addition, the algorithm appears to solve the issue of variations caused by various lighting schemes, viewing direction, shadows, facial expressions, and poses. The researchers were also excited to discover that the algorithm allows for faces to be recognized despite facial hair growth, glasses, and certain other items sometimes employed as disguises.