Dr. Asit Kumar Dutta
Title: Frequency domain correlation filters for object and face recognition
Abstract: While there have been varying and significant levels of performance achieved through the use of spatial 2D image data, the use of a frequency domain representation sometimes achieves better performance for many recognition tasks. The use of the Fourier transforms allow to quickly and easily obtain raw frequency data which are significantly more discriminating (after appropriate data manipulation) than the raw spatial data, from which it is derived. In the majority of cases, correlation filters are used to achieve desired performances due to several advantages, such as 1) it has built-in shift invariance, 2) correlation filters are based on integration operation and thus offer graceful degradation of any impairment to the test object image, 3) correlation filters can be designed to exhibit attributes such as noise tolerance and high ability for discrimination and 4) finally design of correlation filter is derived from closed form expressions and thus physically realizable.
Frequency domain face recognition techniques are executed by cross correlating the Fourier transform of test object image with a synthesized template or filter, generated from the Fourier transform of training object images. The processing results in a correlation output via an inverse Fourier transform and the correlation plane is searched for peak where a perfect match would result in a sharp correlation peak. The relative height of the peak is analyzed by to determine whether the test object is recognized or not and is generally measured by a metric, called peak -to-sidelobe ratio.
The paper describes the techniques of designing correlation filters for recognition of few objects and also for recognition of frontal face images. Practical implementation is included using electronic and photonic technologies. Test results are included.
- Research Areas: Applied Physics, Applied Optics and Photonics