The 2D Fast Fourier Transform (FFT) decomposes an image into its spatial frequency components. In simple terms, it reveals how much of each "pattern" (from fine details to broad gradients) exists in the image. The result is a complex-valued matrix where each point represents a specific spatial frequency — low frequencies (broad shapes, smooth gradients) appear near the center, while high frequencies (sharp edges, fine textures, noise) appear toward the edges.

The magnitude spectrum shows the amplitude (strength) of each frequency component. It tells you how much of each frequency exists in the image, ignoring phase information. Bright spots in the magnitude spectrum indicate dominant frequencies. The DC component (center pixel) represents the average brightness of the entire image. Periodic patterns like stripes produce distinct peaks, while random noise appears as a diffuse background.

The phase spectrum encodes the position information of features in the image. While the magnitude tells you which frequencies are present, the phase tells you where they are located relative to each other. Phase is crucial for reconstructing the original image — if you swap the phase spectra of two images while keeping their magnitudes, the resulting images will look more like the source of the phase data. This demonstrates that phase carries essential structural information about edges, shapes, and object locations.

By default, the FFT output places the zero-frequency (DC) component at the top-left corner, with frequencies increasing toward the bottom-right. We apply fftshift to rearrange the quadrants so that the DC component sits at the center, low frequencies surround it, and high frequencies radiate outward. This centered view is much more intuitive — it resembles how we naturally think about frequency space, with the "heartbeat" of the image at the center and progressively finer details toward the periphery.

The raw magnitude values span an enormous dynamic range — the DC component can be millions of times brighter than high-frequency components. Without logarithmic compression, the spectrum would appear as a single bright dot at the center with everything else nearly black. Applying log(1 + magnitude) compresses this range, making both strong and weak frequency components visible simultaneously. This reveals the rich structure of the frequency domain that would otherwise be hidden.

The frequency domain reveals: (1) Periodic patterns (stripes, grids) appear as sharp peaks — their position indicates orientation and spacing. (2) Sharp edges produce energy across many high frequencies in a characteristic cross pattern. (3) Noise manifests as scattered high-frequency energy. (4) Blur removes high frequencies, so a blurry image has a spectrum concentrated near the center. (5) Image compression artifacts often leave telltale signatures in the frequency domain. Engineers use FFT for image filtering, deconvolution, pattern analysis, and quality assessment.

The phase spectrum often appears noisy because phase values are essentially pseudorandom for many real-world images — they encode fine positional information that doesn't follow a smooth pattern. However, this "noise" is highly structured and contains critical information. We use a cyclic colormap (rainbow/HSL) for phase because phase is periodic (−π and +π represent the same angle). This makes the phase spectrum look like colorful static, but every pixel carries precise geometric information about the image structure.