Understanding Fourier Spectrum Properties in Images

While Fourier transforms play a fundamental role in one-dimensional signal processing, they application in image processing presents unique challenges. The primary difference lies in human perception: audio signals often appear chaotic in their raw form, requiring frequency analysis to reveal their structure, whereas images convey information directly through visual patterns. This makes the interpretation of image spectra more complex.

Periodicity in 2D Fourier Transforms

The discrete Fourier transform exhibits periodicity in both spatial and frequency domains along both X and Y directions. This periodic extension continues infinitely in both domains.

% Create periodic image pattern
baseImage = im2double(imresize(imread('cameraman.tif'), 2));
periodicPattern = repmat(baseImage, 3, 3);

% Compute and visualize spectrum
baseSpectrum = log(abs(fftshift(fft2(baseImage))) + 1;
periodicSpectrum = repmat(baseSpectrum, 3, 3);

figure;
imshowpair(periodicPattern, periodicSpectrum, 'montage');

To facilitate frequency domain analysis, spectrum centering is commonly performed before transformation. This involves multiplying the original matrix by (-1)^(x+y) to shift the DC component to the center.

% Compare centered vs non-centered spectra
originalImage = im2double(imresize(imread('cameraman.tif'), 2));
nonCentered = log(abs(fft2(originalImage)) + 1;
centered = log(abs(fftshift(fft2(originalImage))) + 1;

figure;
imshowpair(nonCentered, centered, 'montage');

Frequency Distribution Patterns

After spectrum centering, the brightest central point represents the DC component (lowest frequency). Higher frequencies appear further from the center, with the corners and axis extremities containing the highest frequency components. Uncentered spectra show the opposite pattern.

% Generate sine waves at different frequencies
signalLength = 512;
samplingRates = [150, 500, 1000];
windowFunction = hanning(signalLength) * hanning(signalLength)';

for i = 1:3
    timeIndex = linspace(0, samplingRates(i), signalLength);
    sineWave1D = sin(timeIndex);
    sineWave2D = repmat(sineWave1D, [signalLength, 1]);
    windowedSine = sineWave2D .* windowFunction;
    
    % Compute and display spectrum
    spectrum = log(abs(fftshift(fft2(windowedSine))) + 1;
    subplot(2, 3, i+3), imshow(spectrum, []);
end

Energy Distribution in Frequency Domain

The DC component contains the majority of spectral energy. Approximately 85% of total energy resides within the central region, 93% within a moderate radius, and nearly 99% within the outer boundaries of the spectrum.

Spatial-Frequency Direction Correlation

Horizontal periodic variations in the spatial domain manifest along the horizontal axis in the frequency spectrum, while vertical variations appear along the vertical axis. Diagonal patterns produce corresponding diagonal spectral components.

% Demonstrate directional correlation
imageSize = 512;

% Create horizontal pattern
horizontalPattern = zeros(imageSize);
horizontalPattern(1:imageSize/2, :) = 1;

% Create vertical pattern  
verticalPattern = zeros(imageSize);
verticalPattern(:, imageSize/2:end) = 1;

% Compute spectra
horizontalSpectrum = log(abs(fftshift(fft2(horizontalPattern))) + 1;
verticalSpectrum = log(abs(fftshift(fft2(verticalPattern))) + 1;

figure;
subplot(2,2,1), imshow(horizontalPattern), title('Horizontal Pattern');
subplot(2,2,2), imshow(verticalPattern), title('Vertical Pattern');
subplot(2,2,3), imshow(horizontalSpectrum, []), title('Horizontal Spectrum');
subplot(2,2,4), imshow(verticalSpectrum, []), title('Vertical Spectrum');