- Biomedical Optics & Medical Imaging
- Defense & Security
- Electronic Imaging & Signal Processing
- Illumination & Displays
- Lasers & Sources
- Micro/Nano Lithography
- Optical Design & Engineering
- Optoelectronics & Communications
- Remote Sensing
- Sensing & Measurement
- Solar & Alternative Energy
- Sign up for Newsroom E-Alerts
- Information for:
Electronic Imaging & Signal Processing
Highly accurate and robust face recognition based on an optical parallel correlator
The fast face recognition optical correlator, applied to a temporal sequence of moving images, has an improved recognition rate and shows promise for applications requiring very high throughput.
1 March 2006, SPIE Newsroom. DOI: 10.1117/2.1200602.0053
In the post 9/11 era, there are increased security demands for accurate personal identification based on individual biometric characteristics. Commercial products using fingerprints, facial images, vein patterns, voiceprints, etc., have been developed and put into service, but none have met all requirements for accuracy, speed, flexibility, and convenience. Because facial images can be captured from a distance without any physical contact, face recognition technology has been drawing increased attention for its potential in this area.
A face recognition system processes images of faces that form a complicated three-dimensional structure. To achieve high recognition accuracy, it is necessary to prepare some sort of dictionary that accommodates various situations. Increasing the size of the database increases the operation time, so attempts have been made to reduce the time of calculation by using the eigen-face method, for example. People can generally wait a maximum of three seconds for their identification to be verified, and current processing time is still too long to handle a large-scale database of 10,000 faces or more accurately and robustly.
We previously developed a face recognition system with an operational speed greater than 4000faces/s.1,2 Our fast face recognition optical correlator (FARCO) is based on the Vanderlugt correlator, combined with high-speed display devices and four-channel processing. We also developed an algorithm for a simple filter by optimizing the calculation algorithm, the quantization digits, and the carrier spatial frequency. This correlation filter is more accurate than the classical correlation. In preliminary experiments on a 1-to-N identification basis, FARCO achieved low error rates of 1% false match rate (FMR) and 2.3% false non-match rate (FNMR). Exploiting the high-speed data-processing capability, significantly more robust recognition can be achieved under various conditions by storing multiple images for each registered person.
Now we have applied this system to a temporal sequence of moving images. The multiplexed database is extracted from video data and contains various images taken from different angles. Using the software,3 we were able to detect four points in a facial image (the positions of the eyes and nostrils) and plot the coordinates at a speed of 100ms/face. The size of the extracted image was normalized to 128×128 pixels by the center of gravity. For input facial images taken at an angle, we used affine transformation to adjust the image and normalization, fixing on the position of the eyes. Edge enhancing with a Sobel filter and binarizing (defining the white area as 20%) equalized the volume of transmitted light in the image. A matched filter was calculated by pre-processed images.
The algorithm for video recognition is shown in Figure 1. The database that registers N people contains N× M images, where M is the number of times a single person is multiplexed. The sequences of L input images are taken from the video camera, and for each input image taken, correlation values are calculated. The highest correlation value is chosen from among the M images of a single registrant. The normalized highest correlation values are averaged over all the N registrants to derive the so-called comparison value. The lowest of the comparison values for a sequence of L input images is compared to the pre-defined threshold to judge whether or not the person in the input images is one of the registrants.
Figure 1. The face recognition algorithm for the FARCO employing a temporal image sequence.
We carried out simulation experiments on the FARCO containing 30 registrants and 30 non-registrants during a period of four weeks. The facial images taken during the first week were used as the database to which the images in the following weeks were compared. Recognition results are shown in Figure 2, where the y-axis represents the recognition rate and the x-axis the number of input facial images. In the cases where 40 multiplexed images and 20 multiplexed input facial images were applied, a high recognition rate of 99.2% was realized. In this experiment, increasing the number of database images M resulted in a higher recognition rate than increasing the number of input images L.
Figure 2. Recognition results.
Our experiments confirm that temporal sequential images functioned effectively as part of the system. This system has promise for a variety of applications where a large number of images have to be handled at high speed, such as security and medicine. We are developing the image search engine that integrates a holographic memory and the optical correlation technology used in FARCO to achieve correlation times of less than 10μs/frame.
Eriko Watanabe and Kashiko Kodate
Faculty of Science, Japan Women's University
Eriko Watanabe obtained her MS and PhD degrees in science in 2002 and in 2005 from the Japan Women's University. She is currently a JSPS Postdoctoral Fellow at Japan Women's University, Department of Mathematical and Physical Science. Her current research interests include optical information processing, pattern recognition—such as face recognition—and holography.
Kashiko Kodate was graduated from Japan Women's University, Tokyo, Japan in 1963 and acquired Ph.D.degree in electronic engineering from the University of Tokyo in 1981. She is currently a professor at Japan Women's University, Department of Mathematical and Physical Science. She has been engaged in researches on diffractive optics and its application to optical information systems, such as facial recognition, arrayed waveguid grating, Talbot array illuminator and optical interconnection.