Research Summary
Face recognition is a widely popular research area made more popular by the recent terrorist activities. I did some initial work in Face recognition and here are some of my works in this field
Face Recognition Using Mahalanobis Distance: The eigen face approach is a widely popular approach which uses the Euclidian distance measure for calculation of similarity. In this paper we use the Mahalanobis Distance measure work as the distance measure for face recognition. ON the ORL database it gave a recognition accuracy of 98.56%.
A Framework for Performance Evaluation of Face Recognition Algorithms
Most face recognition algorithms depend upon some type of similarity measure
to retrieve candidate face images from a database. Many different similarity
measures can be used. However, since human beings seem to be very good at
recognizing faces, it seemed to us that it would be worthwhile to study
what facial characteristics are most salient to humans, as they evaluate
subjective similarity between faces.But the currently available face databases
are not adequate to perform a comprehensive performance evaluation of face
analysis and recognition systems. Specifically, the problem of evaluating
a face recognition system’s tolerance of variations in environmental
parameters (such as pose angle, illumination angle, and illumination color)
can only be done with a comprehensive set of face images that widely populate
the multidimensional space defined by these parameter values. Such a set
of face images can only be generated by methodical control of the environmental
parameter values as the images are captured, and by subsequent use of those
values used to annotate the images in the resulting database.
Unfortunately, most of the currently available sets of face images do not
widely populate this multidimensional environmental parameter space. In
cases where variations in these environmental parameters were used during
the capture of images they were often not methodically controlled, accurately
measured, or used to annotate the images within the resulting image set.
Testing a face recognition algorithm with a set of images that includes
only a few qualitative pose angels (such as frontal, profile and 3/4 view)
does not establish how many degrees of pose angle variation the algorithm
can tolerate. Such coarsely quantized pose angles are also of limited value
in evaluating incremental improvements in the algorithm’s tolerance
of pose angle variations.
In this project, we develop a methodology to create an annotated face database
employing a novel set of apparatus for rapid capture of images.
Figure 1. Face Setup
The figure above shows the experimental set up we built to capture faces. The subject sits in the middle while the camera rotates around him on a concentric disc. The video file is taken and at every 10 degrees a counter is incremented using light sensitive resistors and then recorder. using this we can get a -90 degrees to +90 degrees shot in seconds.
To simulate light angle variations the camera is kept fixed while light revolves around the subject.
Here are some results
Figure 2. Pose Variation
Figure 3. Light Variations
The idea was to have controlled setup to be able to measure light angle and pose angle variations. Any algorithm can test itself against these images and videos and pin point as to what its accuracy is in terms of some common elements. We have modified this setup for object capture as in the visio-haptic setup.
ASU Main Website