Q6Digital Image Processing

(a) Noise Estimation Parameters

Before designing an appropriate restoration filter, it is necessary to estimate the characteristics of the noise corrupting an image. Several parameters and techniques are used for this purpose. The mean and variance of the noise are typically estimated by examining a small region of the image that is known or assumed to be of uniform, constant intensity in the original uncorrupted scene, such as a patch of flat background or sky. Since any variation in pixel intensity within such a strip-shaped or patch-shaped region can only be attributed to noise (as the true underlying signal is constant there), the sample mean and sample variance of the pixel values in that patch directly give an estimate of the noise's mean and variance, which in turn indicates the strength and characteristics of the noise process (e.g., whether it resembles Gaussian noise with a particular standard deviation).

The signal-to-noise ratio (SNR) is another key estimation parameter, defined as the ratio of signal power (or variance) to noise power (or variance); a low SNR indicates heavy noise corruption relative to the underlying image content, and this ratio is often used directly as an input parameter to restoration filters such as the Wiener filter. The image's histogram can also be used for noise estimation: examining the shape of the histogram in a region of known constant intensity reveals the shape of the noise's probability density function (for instance, a Gaussian-shaped histogram bump indicates Gaussian noise, while a histogram with two sharp isolated spikes near the extremes indicates salt-and-pepper impulse noise). For periodic noise, such as interference introduced by faulty electronic acquisition equipment, the periodogram or Fourier spectrum of the image is examined instead of the spatial histogram; periodic noise manifests as distinct, isolated bright impulses or spikes at specific frequency coordinates in the Fourier spectrum (away from the origin), and identifying the locations and strengths of these spikes allows the noise's period and amplitude to be estimated directly from the frequency domain.

These estimated noise parameters play a crucial role in selecting the appropriate restoration approach: for example, a well-estimated noise variance or SNR is a required input parameter for constrained least-squares or Wiener filtering, and knowledge of whether the noise is impulsive versus additive Gaussian versus periodic in nature guides the choice between spatial filters (like median filtering for impulse noise) versus frequency-domain notch filtering (for periodic noise) versus statistical restoration filters (for additive random noise).

(b) Inverse Filtering and Wiener Filtering

Inverse filtering is the most direct approach to image restoration when a degradation function H(u,v) is known. If the degraded image's Fourier transform is G(u,v) = H(u,v)F(u,v) + N(u,v), where F(u,v) is the original image's transform and N(u,v) is additive noise, inverse filtering estimates the original image by simply dividing the observed degraded transform by the degradation function.

This works reasonably well when the degradation function H(u,v) is accurately known and the noise term N(u,v) is negligible, since substituting G(u,v) = H(u,v)F(u,v) gives F_hat(u,v) = F(u,v) exactly. However, inverse filtering has a severe practical limitation: because it divides by H(u,v), at any frequency where H(u,v) has a very small magnitude (or is exactly zero, which commonly occurs at higher frequencies for typical blur degradation functions), the corresponding noise component N(u,v)/H(u,v) becomes extremely large, since it is being divided by a near-zero value. This causes the noise to be massively amplified in the restored image, often to the point of completely overwhelming the recovered signal and making the inverse-filtered result visually far worse than the degraded input, especially when any appreciable noise is present in the degraded image.

Wiener filtering addresses this critical weakness by incorporating an explicit term for the noise-to-signal power spectral density ratio, deriving the filter that minimizes the mean square error between the estimated and the true original image, rather than simply undoing the degradation function.

Here, S_eta(u,v) is the power spectrum of the noise and S_f(u,v) is the power spectrum of the undegraded original image, so the ratio S_eta(u,v)/S_f(u,v) represents the noise-to-signal power ratio at each frequency. When this ratio is zero (i.e., no noise present), the Wiener filter formula reduces exactly to the plain inverse filter 1/H(u,v), recovering the ideal case. However, when noise power is significant relative to signal power at a given frequency (which is precisely where H(u,v) tends to be small, i.e., the same problematic frequencies for inverse filtering), the added ratio term prevents the overall filter gain from blowing up, instead smoothly rolling off the filter's response at those frequencies. This makes the Wiener filter far more robust and noise-tolerant than plain inverse filtering, at the modest cost of requiring an estimate of the noise-to-signal power ratio (which can be obtained using the noise estimation techniques described above), and it is for this reason widely preferred over simple inverse filtering whenever the degraded image contains non-negligible noise.

In many practical situations, the exact power spectra S_eta(u,v) and S_f(u,v) required by the full Wiener filter formula are not individually known, so a simplified variant is commonly used in which the noise-to-signal ratio term is replaced by a single constant K (an approximate average noise-to-signal ratio), giving the widely used approximate form F_hat(u,v) = [ |H(u,v)|^2 / (|H(u,v)|^2 + K) ] x [G(u,v)/H(u,v)]. This constant-K approximation sacrifices some restoration accuracy compared to using the true, frequency-dependent power spectra, but is far easier to apply in practice since K can simply be tuned experimentally (by visual inspection of the restored result) to balance the trade-off between sharper detail recovery (favored by a smaller K, behaving more like plain inverse filtering) and stronger noise suppression (favored by a larger K, behaving more like a low-pass smoothing filter).

Both inverse filtering and Wiener filtering require the degradation function H(u,v) to be known or estimated beforehand, which in practice is obtained either through direct measurement of the imaging system (for example, characterizing the point spread function of a lens or motion blur pattern), through experimentation, or through mathematical modeling of the physical degradation process (such as modeling atmospheric turbulence blur or known linear motion blur with a specific direction and distance). Once H(u,v) is estimated by any of these means, Wiener filtering is generally preferred over plain inverse filtering as the practical restoration method of choice whenever the degraded image is known or suspected to contain non-negligible noise, which is the case in nearly all real acquisition and transmission scenarios.