Better Face Recognition
A new algorithm improves automated recognition of faces in low-resolution images.
Source: “Recognition of Low-Resolution Faces Using Multiple Still Images and Multiple Cameras”
Pablo Hennings-Yeomans et al.
IEEE International Conference on Biometrics: Theory, Applications, and Systems, September 29-October 1, 2008, Crystal City, VA
Results: Researchers at Carnegie Mellon University and Microsoft Research have built a system that improves automated recognition of faces in low-resolution images.
Why it matters: Low-resolution images from surveillance and traffic cameras, cell-phone cameras, and webcams aren’t much use for automatic face recognition, because they lack fine detail. The new system, however, can yield accurate matches from low-quality images. It could be used to search for specific faces on websites, and law-enforcement officials could use it to find suspects in surveillance videos.
Methods: Face recognition systems are usually trained on databases that include many high-resolution images of faces. That teaches them a technique called feature extraction: they learn to associate patterns of pixels with physical traits, such as a particular slant of the eyes. This training, however, doesn’t equip the systems to handle low-resolution images very well. Existing algorithms can increase images’ resolution–adding pixels to smooth out curves, for example. But while the results look better to human beings, the process can cause distortions that lead to errors in automated face recognition. The researchers developed algorithms that improve resolution in ways that take into account the requirements of feature extraction, increasing the accuracy of face identification.
Next steps: Face recognition systems need further improvements to correctly identify images taken from unusual angles. The researchers will also investigate other applications of image recognition–in biomedical imaging, for instance.