Image filters

Filtering is a discipline that crosses physics, electronics, and computation. An interference filter in an optical path passes a band of wavelengths and rejects the rest — separating signal from contamination in the electromagnetic spectrum. An electronic filter in a circuit passes a band of frequencies and rejects the rest — separating signal from contamination in the time domain. An image filter does the same job in the spatial domain — separating the spatial signal you want to keep from the spatial signal you don't.

The shared principle, in every case: signal and contamination rarely occupy the same place in the relevant domain. Real biology — the cells, the structures, the boundaries — lives at one set of spatial frequencies. Sensor read noise lives at another. Bright outliers from hot pixels or contaminating particles live at yet another. Slowly-varying autofluorescence lives at a fourth. The right filter for an image is the one whose pass-band — whether explicit (a Gauss filter passes low spatial frequencies) or implicit (a median filter passes neighborhoods that aren't dominated by outliers) — matches the spatial signal you want to surface.

The contamination isn't always uniform. A steady noise floor across the field is one case; a gradient that varies smoothly across the image is another; a localized region of unusual contamination is a third. Each suggests a different filter, sometimes a chain of filters, sometimes position-aware filtering. Recognizing the character of the contamination — its scale, its uniformity, whether it's random or structured — is the move that determines the right filter, often more than the filter's name.

The fastest path through the filter family is to start from the structure you want to surface. A few common cases:

Isolated extreme pixels alongside the signal you want: Median filterLoading... — sorts the neighborhood and picks the typical value, gracefully ignoring outliers by construction. The elegant move when individual pixels carry signal you don't need to keep.

Broad, low-amplitude variation across the field: Gauss filterLoading... — averages the neighborhood with a smooth weight profile. Reduces variation while preserving the larger features that defined the image.

Smoothing across regions while keeping edges crisp: Bilateral filterLoading... — weights neighbors by both spatial distance and intensity similarity, so smoothing happens within regions but stops at boundaries. The right move when the transitions are part of the structure you care about.

Surfacing edges (cell boundaries, vasculature, structural transitions): SobelLoading... for gradient magnitude, CannyLoading... for clean thin edge lines, LaplaceLoading... for second-derivative responses. Each highlights a different kind of intensity change.

Surfacing membranes specifically: Membrane detectorLoading... — tuned to the rapid double-variation pattern that membranes produce. The kernel knows what it's looking for.

Sharpening for emphasis: Sharp filterLoading... or Unsharp filterLoading.... Sharp filters lift detail; consider their effect on quantitative analysis where the original intensities are the measurement.

Have a specific kernel in mind: Custom filterLoading... — write the weights yourself. The clearest path when the textbook entry isn't the right one.

The engines referenceLoading... has the full per-engine detail. This page is the map of the territory.