2D Fourier transform for images in Signal Processing - Time & Space Complexity

Time Complexity: 2D Fourier transform for images

O(M^2 * N^2)

Understanding Time Complexity

We want to understand how the time to compute a 2D Fourier transform changes as the image size grows.

How does the work needed grow when the image gets bigger?

Scenario Under Consideration

Analyze the time complexity of the following code snippet.

for u in range(M):
    for v in range(N):
        sum = 0
        for x in range(M):
            for y in range(N):
                sum += image[x][y] * exp(-2j * pi * ((u * x / M) + (v * y / N)))
        F[u][v] = sum
    

This code computes the 2D Fourier transform of an M by N image by summing over all pixels for each frequency component.

Identify Repeating Operations

Identify the loops, recursion, array traversals that repeat.

• Primary operation: Nested loops over image pixels and frequency components.
• How many times: The inner sum runs M*N times for each of the M*N frequency points, totaling M*N * M*N operations.

How Execution Grows With Input

As the image size grows, the number of operations grows very fast because we do a full sum for each frequency point.

Input Size (M=N=n)	Approx. Operations
10	10 * 10 * 10 * 10 = 10,000
100	100 * 100 * 100 * 100 = 100,000,000
1000	1000 * 1000 * 1000 * 1000 = 1,000,000,000,000

Pattern observation: The operations grow by the square of the number of pixels squared, which means the time grows very quickly as the image gets bigger.

Final Time Complexity

Time Complexity: O(M^2 * N^2)

This means the time needed grows very fast, roughly with the square of the total number of pixels.

Common Mistake

[X] Wrong: "The 2D Fourier transform runs in time proportional to the number of pixels (M*N)."

[OK] Correct: The code does a sum over all pixels for each frequency point, so the work multiplies, making it much slower than just M*N.

Interview Connect

Understanding how the 2D Fourier transform scales helps you explain performance in image processing tasks and shows you can analyze nested loops clearly.

Self-Check

"What if we used a fast algorithm like FFT for the 2D transform? How would the time complexity change?"

Practice

1. What does the 2D Fourier transform of an image primarily represent?

easy

A. The frequency components of the image

B. The color distribution of the image

C. The brightness levels of the image

D. The pixel coordinates of the image

Solution

1. Step 1: Understand the purpose of 2D Fourier transform

   The 2D Fourier transform converts spatial data into frequency data, showing how much of each frequency is present.

2. Step 2: Relate to image analysis

   In images, this means representing patterns and textures as frequency components rather than pixel values.

3. Final Answer:

   The frequency components of the image -> Option A

4. Quick Check:

   2D Fourier transform = frequency components [OK]

Hint: 2D Fourier transform shows frequencies, not colors or pixels [OK]

Common Mistakes:

• Confusing frequency with color or brightness
• Thinking it shows pixel positions
• Assuming it changes image colors

2. Which Python function is used to compute the 2D Fourier transform of an image array?

easy

A. np.fft.fft

B. np.fft.fft2

C. np.fft.ifft2

D. np.fft.fftshift

Solution

1. Step 1: Identify the function for 2D Fourier transform

   In numpy, fft2 computes the 2D Fourier transform of a 2D array (like an image).

2. Step 2: Differentiate from other functions

   fft is 1D, ifft2 is inverse transform, and fftshift centers the zero frequency.

3. Final Answer:

   np.fft.fft2 -> Option B

4. Quick Check:

   2D Fourier transform function = fft2 [OK]

Hint: Use fft2 for 2D transform, fftshift only re-centers [OK]

Common Mistakes:

• Using fft instead of fft2 for 2D data
• Confusing fft2 with inverse ifft2
• Using fftshift to compute transform

3. Given the code below, what is the shape of F_shifted?

import numpy as np
image = np.random.rand(64, 64)
F = np.fft.fft2(image)
F_shifted = np.fft.fftshift(F)

medium

A. (64, 64)

B. (32, 32)

C. (128, 128)

D. (64,)

Solution

1. Step 1: Check input image shape

   The input image has shape (64, 64), a 2D array with 64 rows and 64 columns.

2. Step 2: Understand fft2 and fftshift output shapes

   The 2D Fourier transform fft2 returns an array of the same shape as input. fftshift only reorders the frequencies but does not change shape.

3. Final Answer:

   (64, 64) -> Option A

4. Quick Check:

   fft2 and fftshift keep shape same as input [OK]

Hint: fft2 and fftshift keep the original image shape [OK]

Common Mistakes:

• Assuming fftshift changes shape
• Confusing shape with half or double size
• Thinking fft2 flattens the array

4. Identify the error in this code snippet for computing and displaying the magnitude spectrum of an image's 2D Fourier transform:

import numpy as np
import matplotlib.pyplot as plt
image = np.random.rand(128, 128)
F = np.fft.fft2(image)
magnitude = np.abs(F)
plt.imshow(magnitude, cmap='gray')
plt.show()

medium

A. Using np.abs instead of np.angle

B. Image size should be square

C. Missing fftshift to center zero frequency

D. plt.imshow requires RGB image

Solution

1. Step 1: Understand the purpose of fftshift

   After fft2, the zero frequency is at the corners. fftshift centers it for better visualization.

2. Step 2: Check code for fftshift usage

   The code computes magnitude but does not apply fftshift, so the spectrum will look off-centered.

3. Final Answer:

   Missing fftshift to center zero frequency -> Option C

4. Quick Check:

   fftshift centers spectrum for display [OK]

Hint: Always use fftshift before displaying magnitude spectrum [OK]

Common Mistakes:

• Skipping fftshift causes off-center spectrum
• Confusing magnitude with phase
• Thinking image must be RGB for imshow

5. You want to remove high-frequency noise from a grayscale image using its 2D Fourier transform. Which approach is best?

hard

A. Set low-frequency components to zero and inverse transform

B. Apply fftshift and keep all frequencies unchanged

C. Use np.abs on the transform and display it

D. Set high-frequency components to zero and inverse transform

Solution

1. Step 1: Identify noise frequency type

   High-frequency components correspond to noise and sharp changes; removing them smooths the image.

2. Step 2: Apply filtering in frequency domain

   Setting high-frequency components to zero removes noise; then inverse transform recovers the cleaned image.

3. Step 3: Understand why other options fail

   Setting low frequencies to zero removes main image info; fftshift only centers frequencies; np.abs just shows magnitude, no filtering.

4. Final Answer:

   Set high-frequency components to zero and inverse transform -> Option D

5. Quick Check:

   Remove noise by zeroing high frequencies [OK]

Hint: Zero high frequencies to remove noise, then inverse transform [OK]

Common Mistakes:

• Removing low frequencies loses main image details
• Thinking fftshift filters noise
• Confusing magnitude display with filtering