Frequency Domain Enhancement: Filters, Transforms & Real-World Apps

TAKEAWAYS:

• Frequency domain separates an image into sine/cosine components (frequencies) for targeted enhancement, unlike spatial-domain methods that operate directly on pixels.
• Low-pass filters smooth images by attenuating high frequencies (e.g., blurring noise), while high-pass filters sharpen edges by amplifying high frequencies.
• Fourier Transform (FT) converts spatial-domain images to frequency-domain spectra, enabling frequency-specific operations (e.g., removing periodic noise).
• Ideal vs. Butterworth filters trade off sharpness and ripple effects: Butterworth filters are smoother but less aggressive in frequency cutoff.
• Homomorphic filtering combines low/high-pass filters to simultaneously enhance illumination and detail in images (e.g., medical X-rays).
• Real-world impact: From Khalti’s OCR (cleaning scanned receipts) to NTC’s traffic cameras (reducing blur in license plate recognition), frequency-domain techniques are invisible but critical in modern systems.

1. Why Frequency Domain? Spatial vs. Frequency Domains

1. Spatial domain: Direct pixel manipulation (e.g., contrast stretching, convolution).
2. Frequency domain: Representing images as combinations of sine/cosine waves (frequencies), enabling selective enhancement.

graph LR
  A["Spatial Domain"] -->|"Direct pixel operations"| B["Contrast Stretching, Convolution"]
  C["Frequency Domain"] -->|"Frequency-based operations"| D["Fourier Transform, Filtering"]
  A -->|"2D DFT"| C
  C -->|"Inverse 2D DFT"| A

Key Idea:

• Low frequencies = smooth regions (e.g., sky in a photo).
• High frequencies = edges/textures (e.g., tree branches).
• Frequency-domain filters let you amplify/dampen specific frequencies without affecting others.

2. The Fourier Transform: From Pixels to Frequencies

The 2D Discrete Fourier Transform (DFT) converts an image $f(x,y)$ into its frequency spectrum $F(u,v)$: $$F(u,v)=\sum _{x=0}^{M-1}\sum _{y=0}^{N-1}f(x,y)\cdot e^{-j2\pi \left(\frac{ux}{M}+\frac{vy}{N}\right)}$$

• Magnitude spectrum $|F(u,v)|$ shows energy at each frequency.
• Phase spectrum $\angle F(u,v)$ preserves spatial structure.

Worked Example: 2D DFT of a Simple Image

Given: A $3\times 3$ image (zero-padded to $4\times 4$ for DFT):

1 1 1 0
1 1 1 0
1 1 1 0
0 0 0 0

Steps:

1. Compute $F(u,v)$ for each $u,v\in \{0,1,2,3\}$.
2. For $u=0,v=0$ (DC component): $$F(0,0)=\sum _{x=0}^3\sum _{y=0}^{3}f(x,y)=9$$
3. For $u=1,v=0$: $$F(1,0)=\sum _{x=0}^3\sum _{y=0}^{3}f(x,y)\cdot e^{-j2\pi \frac{x}{4}}=3+3e^{-j\pi /2}+3e^{-j\pi }+3e^{-j3\pi /2}=0$$ (Result: All non-DC terms cancel out due to symmetry.)

Visualization: Interpretation:

• The DC component (9) dominates—this image is mostly uniform.
• High frequencies (edges) are absent, so the spectrum is sparse.

3. Frequency-Domain Filters

Filters modify the Fourier spectrum to enhance specific features.

A. Low-Pass Filters (Smoothing)

• Goal: Remove high-frequency noise (e.g., salt-and-pepper noise).
• Ideal Low-Pass Filter (ILPF): $$H(u,v)=\{\begin{array}{ll}1& \text{if }D(u,v)\le D_0\\ 0& \text{otherwise}\end{array}$$ where $D(u,v)=\sqrt{u^2+v^2}$ is the distance from the origin.

Problem: Sharp cutoff causes ringing artifacts (Gibbs phenomenon). Solution: Butterworth Filter (smoother transition): $$H(u,v)=\frac{1}{1+{\left(\frac{D(u,v)}{D_0}\right)}^{2n}}$$

• $n$: Order (higher $n$ = smoother cutoff).

Worked Example: Butterworth Filter on Noisy Image Given: A $4\times 4$ image with salt-and-pepper noise:

50 52 50 50
50 255 50 50
50 50 50 50
50 50 50 50

Steps:

1. Compute DFT $F(u,v)$.
2. Apply Butterworth filter with $D_0=5$, $n=2$:
   • For $D(u,v)\le 5$, $H(u,v)\approx 1$.
   • For $D(u,v)>5$, $H(u,v)$ tapers to 0.
3. Inverse DFT to get smoothed image.

Before/After:

B. High-Pass Filters (Sharpening)

• Goal: Enhance edges/textures by amplifying high frequencies.
• Ideal High-Pass Filter (IHPF): $$H(u,v)=\{\begin{array}{ll}0& \text{if }D(u,v)\le D_0\\ 1& \text{otherwise}\end{array}$$
• Laplacian Filter in Frequency Domain: Equivalent to subtracting a smoothed version of the image.

Worked Example: High-Pass Filtering for Edge Enhancement Given: A blurred $4\times 4$ image:

100 102 100 100
100 100 100 100
100 100 100 100
100 100 100 100

Steps:

1. Compute DFT.
2. Apply IHPF with $D_0=3$.
3. Inverse DFT to get edge-enhanced image.

Result: (Note: The output is near-zero except at the edge pixel (102), showing edge enhancement.)

C. Band-Pass Filters

• Goal: Isolate specific frequency bands (e.g., removing periodic noise).
• Example: Remove power-line interference (50/60 Hz) from medical images.

4. Homomorphic Filtering: Enhancing Illumination and Reflectance

Problem: Images often have uneven illumination (e.g., shadows in X-rays). Solution: Separate the image into:

• Illumination component (low frequencies).
• Reflectance component (high frequencies).

Homomorphic Filter: $$H(u,v)=\left(\frac{{\gamma }_H}{{\gamma }_L}\right)\cdot \frac{{\gamma }_L^{\alpha }+{\gamma }_H^{\alpha }}{{\gamma }_L^{\alpha }+{\gamma }_H^{\alpha }\cdot {\left(\frac{D(u,v)}{D_0}\right)}^{2n}}$$

• ${\gamma }_L$: Low-frequency gain (boost illumination).
• ${\gamma }_H$: High-frequency gain (enhance detail).
• $\alpha$: Controls the transition sharpness.

Worked Example: Homomorphic Filtering on a Medical X-Ray Given: An X-ray with uneven lighting:

Steps:

1. Compute DFT.
2. Apply homomorphic filter with:
   • ${\gamma }_L=0.5$ (reduce illumination).
   • ${\gamma }_H=2.0$ (enhance edges).
   • $D_0=10$, $n=1$.
3. Inverse DFT to get balanced image.

Before/After:

5. Practical Applications in Nepal and Globally

## In the Real World

1. Khalti’s OCR System:

   • Problem: Scanned receipts often have noise or low contrast.
   • Solution: Frequency-domain filtering (low-pass to smooth, high-pass to sharpen text) before OCR.
   • Example: A receipt with handwritten notes is processed to extract clean text for transaction verification.

2. NTC’s Traffic Camera Surveillance:

   • Problem: License plates are blurred due to motion or low light.
   • Solution: Homomorphic filtering to enhance plate edges while reducing glare from headlights.
   • Example: A car’s plate at night:

3. Nepal Rastra Bank’s Currency Validation:

   • Problem: Counterfeit notes may have subtle frequency artifacts.
   • Solution: Band-pass filters to detect missing high-frequency details (e.g., microprinting).
   • Example: A $1000$ NPR note’s security thread is analyzed in the frequency domain to verify authenticity.

4. Google Photos’ "Enhance" Feature:

   • Problem: Photos taken in low light or with camera shake appear noisy.
   • Solution: Adaptive Butterworth filtering to reduce noise while preserving edges.
   • Example: A blurry selfie becomes sharp without losing detail.

5. Pathao’s Driver App:

   • Problem: GPS coordinates may have high-frequency jitter.
   • Solution: Low-pass filtering to smooth driver location data for accurate route tracking.

6. Comparison Table: Frequency-Domain Techniques

7. Step-by-Step: Frequency-Domain Enhancement Pipeline

flowchart LR
  A["Input Image"] --> B[Compute 2D DFT
  (F("u,v") = DFT{"f(x,y)"})]
  B --> C[Design Filter
  (e.g., Butterworth, Ideal High-Pass)]
  C --> D[Apply Filter
  (H("u,v") * F("u,v"))]
  D --> E["Inverse 2D DFT"]
  E --> F["Enhanced Image"]

Worked Example: End-to-End Filtering Given: A $4\times 4$ image with noise:

50 55 50 50
50 255 50 50
50 50 50 50
50 50 50 50

Task: Remove the noise pixel (255) using a Butterworth low-pass filter ($D_0=2$, $n=1$).

Steps:

1. DFT: Compute $F(u,v)$ for all $u,v$. The noisy pixel contributes high-frequency components.
2. Filter Design: $$H(u,v)=\frac{1}{1+{\left(\frac{D(u,v)}{2}\right)}^2}$$
   • For $D(u,v)=0$: $H=1$.
   • For $D(u,v)=2$: $H=0.2$.
3. Filter Application: Multiply $F(u,v)$ by $H(u,v)$. High frequencies (noise) are attenuated.
4. Inverse DFT: Reconstruct the image. The 255 pixel is reduced to ~50.

Result:

8. Common Pitfalls and Exam Tips

## Exam Tip

1. DFT vs. FFT:

   • Always mention that FFT (Fast Fourier Transform) is used for efficient computation in practice, but exams may ask for the DFT formula.
   • Example: "The DFT converts spatial pixels to frequency coefficients, while FFT computes it in $O(N\log N)$ time."

2. Filter Design:

   • Butterworth is preferred over ideal filters because it avoids ringing artifacts.
   • Memorize the Butterworth formula and its parameters ($D_0$, $n$).

3. Homomorphic Filtering:

   • This is a high-scoring topic. Explain it as combining low-pass (illumination) and high-pass (detail) filters.
   • Example: "Homomorphic filtering is used in medical imaging to separate shadows from actual structures."

4. Numerical Problems:

   • For DFT calculations, zero-pad small images to simplify computation (e.g., $3\times 3$ → $4\times 4$).
   • Example: "Given a $2\times 2$ image, pad it to $4\times 4$ before computing DFT."

5. Real-World Applications:

   • Link theory to Nepali contexts (e.g., "NTC uses frequency-domain filtering to enhance CCTV footage").
   • Example: "A Daraz delivery photo with motion blur can be sharpened using high-pass filtering."

6. Common Mistakes:

   • Forgetting inverse DFT: Always reconstruct the image after filtering.
   • Incorrect filter application: Multiply the filter $H(u,v)$ with the spectrum $F(u,v)$, not the original image.
   • Units: Ensure $D_0$ is in the same units as $D(u,v)$ (usually normalized to image dimensions).

9. Summary Checklist

Before the exam, ensure you can:

• Define DFT and explain its role in frequency-domain processing.
• Derive a simple 2D DFT for a small image (e.g., $2\times 2$).
• Design Butterworth and ideal filters and explain their differences.
• Apply homomorphic filtering to a given image (describe steps).
• List two real-world applications of frequency-domain techniques in Nepal.
• Solve a numerical problem involving DFT and filtering (e.g., noise removal).