nitin@ur-laser.uucp (Nitin Sampat) (07/20/85)
Homomorphic(nonlinear) vs. linear filtering of real world photographic images. I recently worked on this topic for my Master's thesis and I thought I would share it with you guys. What I did : Process real - world images ( amateur photographs ) with different types of transfer functions in the frequency domain trying to effect image enhancement. The images were processed both linearly and nonlinearly(homomorphically) and "quality improvement" was effected. The "improvement" was determined by subjective testing on a group of human observers. The justification for this research comes from an interest by Eastman Kodak to process amateur pictures in real time. Since the human eye has a logarithmic response (nonlinear) one would assume that nonlinear operations would have better subjective improvement than linear filtering. Stockham ( proc IEEE,august 1968 ) first proposed homo. filtering wherein one takes the log of the image before taking its FFT, multiplying the image and the transfer function in the frequency domain, inverse transforming and then exponentiating to get the original signal. In linear processing, the exact same procedure is repeated except you don't take the log before Fourier transforming. What I found : Homomorphic filtering although theoretically an obvious choice, has some practical problems. (I worked with 512 X 512, 8 bit images ona Vax-11/785 ) The final image has very few gray levels and appears very dark. Hence, it requires substantial post processing (contrast stretching, histogram equalization etc.) before an acceptable image could be displayed. I found histogram specificatioa(Frei) and histogram equalization both did a "good job"). Linear filtering does not require any significant post processing. After subjecting the same image to linear and nonlinear filtering and the same post processing(to be consistent), subjective analysis revealed that to the amateur photographer, there is very little difference in the 2 techniques. In general, however, linear filtering was preferred. Conclusion : Use linear filtering to process amateur type images. P.S. Please note that this a brief summary and a lot of factors, conditions, variables are not mentioned. As such, these statements should not be regarded as absolute. If someone is interested in knowing about details or has some specific question, send me mail and I'll be glad to reply. Also, I wrote programs in Fortran - 77 to do things like read and write images, compute power spectrum of image, and ofcourse to effect linear and nonlinear filtering. If you are interested, let me know. Nitin {seismo,allegra}!rochester!ur-laser!nitin