MECH5720 Chap.5 Multidimensional Signals and Fourier Analysis
Multidimensional Signals and Fourier Analysis
Two-dimensional Fourier analysis represents an image as spatial-frequency components with magnitude and phase. It turns convolution into multiplication, exposes directional patterns and clarifies sampling and filtering. Low frequencies describe slowly varying structure, while high frequencies describe rapid variation rather than importance by themselves.
Correct interpretation requires attention to origin placement, axis units, conjugate symmetry, boundary assumptions and the distinction between changing displayed magnitude and changing the underlying complex spectrum. A two-dimensional sinusoid is defined by a frequency vector. Its direction is the direction in which phase changes, so visible constant-phase stripes lie perpendicular to that vector.
A real cosine produces a pair of complex-conjugate impulses at positive and negative frequency. Horizontal intensity variation therefore appears on the horizontal frequency axis while the image displays vertical stripes. Magnitude states how much of a component is present. Phase states where components align in space and is essential to edges and recognisable structure.
A logarithmic magnitude display makes weak coefficients visible but changes their apparent brightness; shifting a display moves zero frequency to the centre but does not alter the transform. Axis spacing must be related to image size and sample spacing before coordinates become cycles per pixel or cycles per millimetre. Convolution supplies the main imaging connection.
If y = h*x for a linear shift-invariant system, then Y = HX. The point-spread function and transfer function are thus two descriptions of one operator. Multiplication by a low-pass mask reduces rapid variation and broadens transitions, while a high-pass operation emphasises changes and can also strengthen noise. Abrupt frequency masks create spatial ringing because narrow support in one domain spreads in the other.
Sampling repeats the continuous spectrum. When replicas overlap, high frequencies fold into the baseband as aliases. The Nyquist limit is a condition on the sampled bandwidth, not a promise that every pattern below it will be measured well by the optics. Anti-alias filtering must operate before discrete sampling. Enlarging an already aliased image changes display density without recovering the original component.
Finite image boundaries need explicit treatment. The discrete transform assumes periodic extension, so opposite edges meet under circular convolution. Zero padding, reflection or other extension choices create different boundary evidence and should be reported. Windowing reduces an abrupt crop at the cost of spreading or attenuating frequencies.
A reliable analysis predicts the spatial effect of every spectral action, checks conjugate symmetry for real images and verifies a filter on known sinusoids before applying it to an unknown scene.
What this chapter covers
- 01
Two-dimensional sinusoids
- 02
Fourier magnitude and phase
- 03
Convolution theorem
- 04
Frequency-domain filtering
- 05
Sampling and spectral replicas
- 06
Directional frequency patterns
Worked example · free
Predict a sinusoid location in the spectrum
- +1The variation depends on x, so its frequency vector lies on the horizontal frequency axis.
- +2A real cosine produces a conjugate pair at fx = plus and minus 0.2 cycles per pixel.
- +1Constant-phase lines are vertical, so the visible stripes are vertical.
- +1The frequency vector is perpendicular to the stripe orientation.
Key terms
- Fourier transform
- A representation of a signal by complex spatial-frequency coefficients.
- Fourier magnitude
- The amplitude of each frequency coefficient.
- Fourier phase
- The spatial alignment information carried by each complex coefficient.
- Convolution theorem
- The rule that spatial convolution becomes frequency-domain multiplication.
- Spectral replica
- A repeated spectrum created by sampling a continuous signal.
Multidimensional Signals and Fourier Analysis FAQ
Why is phase important for images?
Phase carries the relative alignment of frequency components that forms edges and recognisable structure. Preserving magnitude while scrambling phase can destroy scene geometry even though the energy at each frequency is unchanged.
What does convolution become after transformation?
For a linear shift-invariant system, y = h*x becomes Y = HX. This makes filtering and transfer analysis direct, while also exposing instability when the transfer coefficient H is close to zero.
How should a shifted Fourier display be read?
Display shifting places zero frequency at the centre for interpretation; it does not alter the underlying frequency content. Axis labels and sample spacing are still needed to convert displayed position into cycles per pixel or physical frequency.
Exam move
Sketch simple vertical, horizontal and diagonal sinusoids beside their paired Fourier impulses. Rehearse the convolution theorem in both directions and attach each filter to its expected spatial effect. Include phase in explanations, label frequency units and check whether a displayed spectrum has been centred before reading locations.
Generate a prediction before viewing any transform: peak axis, signed pair, distance from zero and stripe orientation. Repeat with a frequency beyond Nyquist and identify the folded location. For filters, draw both the spectral mask and the expected edge response, noting ringing from abrupt transitions. Change the boundary extension and explain any wrap-around artifact.
Finish by describing what magnitude alone cannot reveal and why phase must be retained for structure. Verify one prediction numerically on a small sampled grid and label the discrete frequency bins.