Image Filtering — Blur, Sharpen, Edge-Detect (a 1D Slice)
A digital image is just a 2D array of pixel intensities — and the Fourier transform takes it into spatial-frequency space. Low frequencies = smooth regions. High frequencies = edges & texture. Photoshop's blur, sharpen and edge-detect tools are all this idea. Here's the 1D version so you can see exactly what's happening.
Frequency-domain filtering
Reference: Wikipedia — Digital image processing; Gonzalez, R. C., & Woods, R. E. (2018). Digital Image Processing, 4th ed., Pearson, Ch. 4. Also: GeeksforGeeks — Fourier Transform.
🧪 Try these experiments in order
- Look at the original signal — it's a synthetic "image row" with smooth regions, a sharp edge in the middle, and some fine ripples on the right.
- Pick Profile B from the dropdown — a noisy gradient. The low-pass demoes Photoshop's Gaussian blur.
- Drag cutoff very low. The low-pass becomes a flat blob (only DC survives). The high-pass keeps everything except the average — that's how edges pop out.
- Try cutoff just below the position of the fine ripples. The low-pass now smooths only the high-frequency texture, leaving the edge sharp — this is the principle of selective denoising.
Original "image row" (pixel intensity along x)
Low-pass result — smooth regions kept, fine detail removed (blur)
High-pass result — only edges & rapid changes (edge map)
⚠ Watch out for
- Brick-wall cutoffs (like this demo) produce ringing artefacts near edges — real image processing uses Gaussian or Butterworth filters which roll off gently.
- Cutoff = 1 means "only the DC component" — the low-pass becomes a flat horizontal line, not "the picture but blurry". That's a maximally aggressive blur, not an "image" any more.
✅ Do
Use frequency-domain filtering for global operations on whole images (denoise, deconvolve, periodic-noise removal — e.g. removing scanline patterns from a TV capture).
❌ Don't
Use full-image FFT for tasks that are inherently local (face detection, object detection) — small-kernel spatial convolution is faster and avoids edge-artefact issues.
Where this matters in industry
Photoshop & GIMP (every blur, sharpen, unsharp-mask filter), medical imaging (MRI is literally a measurement in Fourier space — the raw data is called "k-space"), JPEG compression (discrete cosine transform = Fourier cousin), astronomy image deconvolution, fingerprint enhancement, satellite imagery.
🎯 Learning checkpoint
An MRI scanner measures the Fourier transform of the patient's body directly. Why is taking more samples in k-space (Fourier space) equivalent to taking a higher-resolution image?