Q3Digital Image Processing

Question

Q.2. Describe the various types of frequency domain filters. [16]

Answer 1

Introduction to Frequency Domain Filtering

Frequency domain filtering operates on the two-dimensional Fourier transform F(u,v) of an image, multiplying it by a filter transfer function H(u,v) to produce a filtered result G(u,v) = H(u,v)F(u,v), which is then converted back to the spatial domain via the inverse Fourier transform. Low frequencies in F(u,v) correspond to slowly varying regions of the image (smooth backgrounds, overall brightness), while high frequencies correspond to rapid intensity transitions such as edges, fine detail, and noise. Filters are classified as low-pass (attenuating high frequencies, producing smoothing/blurring), high-pass (attenuating low frequencies, producing sharpening/edge enhancement), band-pass (passing a specific range of frequencies), and band-reject or notch filters (removing a specific range of frequencies, useful against periodic interference).

Ideal Low-Pass Filter (ILPF)

The Ideal Low-Pass Filter has a transfer function that is 1 for all frequencies within a cutoff radius D0 from the origin of the frequency plane, and 0 for all frequencies beyond D0.

H (u, v) = {1, 0, D (u, v) \leq D_{0} D (u, v) > D_{0}

Here D(u,v) is the Euclidean distance of point (u,v) from the origin (zero frequency) of the frequency plane. Because this filter has an abrupt, discontinuous cutoff, its inverse Fourier transform (the equivalent spatial-domain convolution kernel) has a sinc-function shape with significant side-lobes. Convolving an image with this kernel introduces visible ringing artifacts (spurious oscillatory patterns, known as the Gibbs phenomenon) around sharp edges in the filtered image, which is the main practical drawback of the ideal filter despite its conceptually simple definition.

Butterworth Low-Pass Filter (BLPF)

The Butterworth Low-Pass Filter of order n provides a smooth transition between the passed and attenuated frequencies rather than an abrupt cutoff.

H (u, v) = \frac{1}{1 + [ D ( u , v ) / D _{0} ] ^{2 n}}

As the order n increases, the transition band narrows and the filter behavior approaches that of the ideal low-pass filter, but for small to moderate n the transition remains gradual, which significantly reduces ringing artifacts compared to the ideal filter, at the cost of a slightly less sharp separation between passed and blocked frequencies. The Butterworth filter therefore offers a practical, tunable trade-off between smoothing strength and ringing suppression via the order parameter n.

Gaussian Low-Pass Filter (GLPF)

The Gaussian Low-Pass Filter has a smooth, bell-shaped transfer function with no discontinuities anywhere.

H (u, v) = e^{- D^{2} (u, v) /2 D_{0}^{2}}

Since the Gaussian function and its Fourier transform are both Gaussian (with no oscillatory side-lobes), the corresponding spatial-domain convolution kernel is also a smooth Gaussian with no negative side-lobes. Consequently, the Gaussian low-pass filter introduces absolutely no ringing artifacts, producing the visually smoothest result among the three, though its transition between passed and blocked frequencies is more gradual than even a low-order Butterworth filter.

High-Pass Filters

The corresponding high-pass filters are obtained by subtracting the low-pass transfer function from 1, since high-pass filtering is the complement of low-pass filtering in the frequency domain.

H_{H P} (u, v) = 1 - H_{L P} (u, v)

Ideal High-Pass Filter (IHPF): H(u,v) = 0 for D(u,v) <= D0 and H(u,v) = 1 for D(u,v) > D0; suffers from ringing artifacts around edges, same cause as the ideal low-pass filter.
Butterworth High-Pass Filter (BHPF): H(u,v) = 1 / (1 + [D0/D(u,v)]^(2n)); gives a smooth, order-controlled transition and greatly reduced ringing compared to the ideal high-pass filter.
Gaussian High-Pass Filter (GHPF): H(u,v) = 1 - exp(-D^2(u,v)/2D0^2); gives the smoothest possible transition and produces no ringing at all, though edge enhancement is more gradual.

Band-Pass, Band-Reject, and Notch Filters

Band-pass filters allow only a specific annular range of frequencies (between an inner and an outer radius in the frequency plane) to pass, attenuating both lower and higher frequencies outside that band; they are useful for isolating texture or periodic structures with a known frequency range. Band-reject filters do the opposite, attenuating a specific range of frequencies while passing all others, and notch filters are a special, spatially localized case of band-reject filters that suppress energy only at specific isolated frequency coordinates (and their symmetric counterparts). Notch filters are particularly valuable for removing periodic noise, such as scan-line interference or moire patterns, which appears as distinct, isolated bright spots in the Fourier spectrum away from the origin; by placing narrow notches exactly at these spot locations, the periodic interference can be removed while leaving the rest of the image spectrum, and hence the image content, largely undisturbed.

Practical Comparison

In summary, all three low-pass/high-pass filter families (ideal, Butterworth, Gaussian) achieve the same basic goal of suppressing or passing frequency ranges, but they differ fundamentally in the sharpness of their transition band. The ideal filter's abrupt cutoff causes the most severe ringing due to the Gibbs phenomenon; the Butterworth filter allows a designer to trade off sharpness versus ringing through the order parameter n, with higher orders approaching ideal-filter behavior and its associated ringing; and the Gaussian filter, having a perfectly smooth transfer function, produces the least ringing but the most gradual frequency separation. The choice among them in practice depends on whether preserving a very sharp frequency cutoff or minimizing visual ringing artifacts is more important for the given application.

The choice of cutoff frequency D0 is equally important as the choice of filter family for all these filter types. A very small D0 (a tightly restrictive low-pass filter, or equivalently a low-pass filter that passes only a small central region of the frequency plane around the origin) removes most of the high-frequency content, producing a heavily blurred, smoothed image with most fine detail and noise both removed; a larger D0 preserves progressively more high-frequency detail while still removing at least the very highest frequencies typically associated with noise. Symmetrically, for high-pass filters, a small D0 blocks only the very lowest frequencies (preserving most of the image's mid and high frequency detail while removing gross average brightness), while a larger D0 removes a wider range of low and mid frequencies, leaving only the very sharpest edges and finest detail, and producing a darker overall result since the average (DC) brightness component is always removed by a pure high-pass filter.

It is also useful to note that all of these frequency-domain filters (ideal, Butterworth, Gaussian, and their high-pass counterparts) are radially symmetric functions of D(u,v), meaning their value depends only on the distance from the origin of the frequency plane and not on direction; this is why they are described purely in terms of D(u,v) rather than separately in terms of u and v. This radial symmetry ensures that these filters treat image content uniformly regardless of the orientation of edges or texture in the spatial domain, in contrast to directional filters that might selectively enhance or suppress edges of a particular orientation. In practical implementation, these frequency-domain filters are computed on a frequency grid of the same dimensions as the (typically zero-padded) image's 2D discrete Fourier transform, multiplied element-wise with the transform of the image, and the result is converted back to the spatial domain using the inverse discrete Fourier transform to obtain the final filtered image.