Filters & Frequencies — Edges, Hybrid Images, and Blending

Most of classical image processing is frequency analysis in disguise. Edge detection is a high-pass filter. Sharpening is the same operation with the original added back. Hybrid images combine one picture’s high frequencies with another’s low frequencies — your brain picks which one to see based on viewing distance. Seamless image blending splits images into frequency bands, blends each band with an appropriately-sized mask, then reconstructs.

Edge Detection

Partial derivatives encode edge orientation; their magnitude encodes edge strength. Gaussian smoothing before differentiation (Derivative-of-Gaussian filters) dramatically reduces noise — a first-order derivative applied to noisy pixels just produces more noise.

Unsharp Masking

Sharpening is artificial, but in a controlled way. The blurred version of an image contains only low frequencies; subtracting it from the original isolates high frequencies. Adding those high frequencies back, scaled by α, amplifies edges without changing overall tonality. Applied here to a Taj Mahal photograph and an Inception film frame.

Hybrid Images

Look closely: flowers and a ladybug. Step back: just a ladybug. Get closer to your screen: just a flower. Hybrid images exploit the fact that low-frequency content dominates perception at distance, while high-frequency content dominates up close. Each image here combines high-frequency components of one photo with low-frequency components of another.

Multiresolution Blending

The “oraple” — half orange, half apple — is the classic demo. Naive cut-and-paste produces a hard visible seam. Laplacian pyramid blending splits each image into frequency bands, applies a mask of appropriate softness to each band (hard cut for high frequencies, soft blur for low), then sums the result. The transition becomes invisible because high-frequency mismatches are localized while low-frequency color differences are gradually interpolated.