A new technique is presented for the restoration of images degraded by a linear, shift-invariant blurring point-spread function in the presence of additive white Gaussian noise. The algorithm uses overlapping variable-size, variable-shape adaptive neighborhoods (ANs) to define stationary regions in the input image and obtains a spectral estimate of the noise in each AN region. This estimate is then used to obtain a spectral estimate of the original undegraded AN region, which is inverse Fourier transformed to obtain the space-domain deblurred AN region. The regions are then combined to form the final restored image. Mathematical derivation and implementation of the adaptive-neighborhood deblurring (AND) filter is discussed, and experimental results are presented with an analysis of the performance of the AND filter as compared to the fixed-neighborhood sectioned deblurring (FNSD) Wiener and power spectrum equalization filters. It is shown that using the AND algorithm for image deblurring enables the identification of relatively stationary regions. This improves the restoration process and produces results
that are superior to those obtained using the FNSD method both
visually and in terms of quantitative error measures.