Q6Digital Image Processing

(a) Order-Statistic Filters and Their Applications

Order-statistic filters are nonlinear spatial filters that operate by ranking (sorting) all pixel intensity values within a defined neighborhood window and selecting a specific ranked value from that sorted sequence as the output. Because the output is chosen from ranked values rather than computed as a weighted sum, these filters handle impulse noise far more effectively than linear filters, while also preserving edges better.

• Median filter: outputs the middle (50th percentile) value of the sorted neighborhood; excellent for removing salt-and-pepper noise since extreme outlier values are pushed to the sorted extremes and excluded from the middle position, while also preserving edges much better than a linear average filter.
• Max filter: outputs the maximum (100th percentile) value of the sorted neighborhood; used to find the brightest points locally and is effective for reducing dark 'pepper' impulse noise, though it brightens and grows bright features overall.
• Min filter: outputs the minimum (0th percentile) value of the sorted neighborhood; used to find the darkest points locally and is effective for reducing bright 'salt' impulse noise, though it darkens and grows dark features overall.
• Midpoint filter: outputs the average of the maximum and minimum values, (max+min)/2, within the neighborhood; works best for randomly distributed noise types such as Gaussian or uniform noise, combining some order-statistic character with an averaging step, though it is less effective specifically against salt-and-pepper impulse noise than the pure median filter.

These filters are applied by sliding the chosen window across every pixel position, sorting the neighborhood values at each position, and outputting the appropriate ranked statistic as the new pixel value, with suitable boundary handling (zero-padding or replication) applied at the image edges. Their suitability depends heavily on the specific type of noise present: median filters are the standard choice for impulse (salt-and-pepper) noise removal in general image restoration and preprocessing pipelines, while max/min filters serve more specialized roles targeting one-sided (only bright or only dark) impulse noise.

(b) Inverse Filtering for Degradation Removal

When an image has been degraded by a known degradation function H(u,v), the observed degraded image's Fourier transform can be modeled as G(u,v) = H(u,v)F(u,v) + N(u,v), where F(u,v) is the Fourier transform of the original, undegraded image and N(u,v) represents additive noise. Inverse filtering attempts to recover an estimate of the original image by dividing the observed degraded transform by the known degradation function:

The overall process is: (1) transform the degraded image into the frequency domain via the Fourier transform to obtain G(u,v), (2) divide G(u,v) by the known or estimated degradation function H(u,v) at every frequency point to obtain the estimate F_hat(u,v), and (3) apply the inverse Fourier transform to F_hat(u,v) to bring the estimated, restored image back into the spatial domain. When H(u,v) is accurately known and the noise term N(u,v) is negligible, this recovers F(u,v) essentially exactly, since G(u,v)/H(u,v) = F(u,v) + N(u,v)/H(u,v) reduces to just F(u,v) when N(u,v) is near zero.

The significant practical limitation of inverse filtering is its severe noise amplification: at frequencies where the degradation function H(u,v) has a small magnitude (or is exactly zero, which commonly happens at higher frequencies for typical blur-type degradations), the noise term N(u,v)/H(u,v) becomes extremely large because it is divided by a near-zero denominator. This causes noise to be massively amplified in the restored image, frequently to the point of completely dominating and obscuring the recovered signal, making the naive inverse-filtered result often visually worse than the original degraded image whenever any appreciable noise is present.

A practical mitigation for this problem is restricted or pseudo-inverse filtering, in which the inverse filtering operation (dividing by H(u,v)) is applied only within a limited radius from the origin of the frequency plane, specifically the region where H(u,v) remains reasonably large in magnitude and therefore the noise amplification stays under control; outside this safe radius, where H(u,v) becomes small and noise amplification would be severe, the filter response is instead set to zero (or smoothly tapered toward zero) rather than being allowed to divide by a near-zero denominator. This restricted approach sacrifices the ability to restore genuine high-frequency detail in the heavily noise-dominated region, but avoids the catastrophic noise blow-up that unrestricted inverse filtering would produce, yielding an overall more visually usable and practically acceptable restoration result.

In many practical restoration systems, the choice of the safe radius for restricted inverse filtering is made by inspecting the magnitude spectrum of the degradation function H(u,v) and identifying the frequency at which |H(u,v)| drops below some small fraction of its maximum (near-origin) value, beyond which noise amplification would dominate. An alternative and generally superior practical approach, whenever an estimate of the noise power spectrum is available, is to replace this heuristic hard radius cutoff with the Wiener filter described in the corresponding order-statistic filter question, since the Wiener filter's noise-to-signal ratio term achieves a similar effect of suppressing the problematic near-zero-H(u,v) high frequency region, but does so smoothly and in a mathematically optimal (minimum mean square error) fashion rather than through an ad-hoc, manually chosen sharp radius cutoff.

A significant practical limitation of the basic (unrestricted) inverse filter is that it attempts to fully invert the degradation function H(u,v) at every frequency, including frequencies where H(u,v) is very small or exactly zero - at such frequencies, dividing the (noisy) degraded image spectrum G(u,v) by this near-zero H(u,v) value produces an extremely large, noise-dominated result, since any small amount of noise present in G(u,v) at that frequency gets amplified by the correspondingly large reciprocal 1/H(u,v) factor, often overwhelming the actual restored image signal at those frequencies with amplified noise. This motivates the practical technique of restricted (pseudo-) inverse filtering, in which the inverse filtering operation F_hat(u,v)=G(u,v)/H(u,v) is applied only within a limited radius of frequencies around the origin (where H(u,v) remains reasonably large and well-behaved), while frequencies beyond this radius (where H(u,v) becomes small and inverse filtering would primarily amplify noise) are instead left unmodified or set to zero, trading off some loss of restoration accuracy at the highest frequencies for substantially improved overall image quality by avoiding catastrophic noise amplification - this restricted approach represents a practical middle ground between the full (but noise-sensitive) inverse filter and the more sophisticated Wiener filter, which explicitly incorporates a noise-to-signal power ratio term to achieve a similar noise-robustness benefit in a more mathematically principled way.