Input image spectral density estimation for real-time adaption of correlation filters

The problem of image noise estimation for improved noise robustness and dis crimination capabilities of optical correlation filters is discussed. Color ed noise is often used in the literature as an approximation to the true no ise spectral density in the input image of a correlator. This conjecture is verified on different kinds of input images, i,e., their power spectral de nsities are fitted to a colored noise model. The quality of the resulting a pproximation is discussed, We then show that incorporating this noise estim ation into optimal trade-oft filters can significantly improve both the dis crimination capabilities and the SNR of the resulting adaptive correlation filter above that of the classical filters for which the noise parameters a re not estimated. although its performance is in general found to be marked ly inferior to that of true nonlinear filtering techniques that are optimal for adaptive image correlation, the proposed adaptive method is attractive in terms of computation time. Possible optical implementations of the prop osed method are also discussed. (C) 1999 Society of Photo-Optical Instrumen tation Engineers. [S0091-286(99)01604-9].