Designing Perceptually-Based Image Filters in the Modulation Domain

Event Date: May 3, 2011
Speaker: Professor Joseph P. Havlicek
Speaker Affiliation: University of Oklahoma
Sponsor: Communications, Networking, Signal & Image Processing
Time: 10:30 AM
Location: MSEE 239
Contact Name: Professor Edward Delp
Contact Phone: (765) 494-1740
Contact Email: ace@purdue.edu
Open To: ACCEPTABLE FOR ECE 694A

 

The theory of LTI signal processing is based on representing signals in terms of sinusoidal functions, all of which are Eigenfunctions of any LTI system. In a filter, each Eigenfunction is multiplied by an Eigenvalue that scales the magnitude and shifts the phase. This leads to the elegant classical theory where one designs the frequency response to achieve a desired signal processing goal. But the human senses do not /perceive/ signals directly in terms of the Eigenfunction representation; rather, the human auditory and vision systems perceive signals in terms of localized nonstationary instantaneous amplitude and frequency. In human hearing, the relationship between the sinusoidal basis functions and the instantaneous quantities is often “close enough.”  Thus, in music reproduction we think of designing a “loudness filter” to boost bass and treble while lightly attenuating midrange. The relationship is often not so intuitive in human vision however.  In this talk, I will present a biologically motivated modulation domain theory for representing images directly in terms of visually important instantaneous amplitude and frequency modulations.  Using this theory, I will demonstrate the design of nonlinear filters that can achieve signal processing goals which are directly related to human visual perception. With this new approach we are able to address perceptually motivated design problems such as “design a filter to remove the bands from Lena’s hat” or “design a filter to change the stripes on Barbara’s pants from vertical to horizontal.”

 

Biography

 

 

Joseph P. Havlicek is the Williams Presidential Professor of Electrical & Computer Engineering at the University of Oklahoma. He received the B.S. and  M.S. degrees at Virginia Tech and the Ph.D. degree at the University of Texas at Austin.  Before joining OU, he was on the scientific staff of the U.S. Naval Research Laboratory and was a software developer at IBM. He is currently an associate editor of IEEE Transactions on Image Processing and IEEE Transactions on Industrial Informatics.  Dr. Havlicek's main research areas include modulation domain signal processing, statistical signal theory, discrete uncertainty measures, and intelligent transportation systems.