Anisotropic filter

Start with the register the Engines Reference speaks in. Anisotropic filter sits in the Basic Operations Module under FiltersLoading.... The reference describes anisotropic diffusion as a way to reduce image noise while preserving significant information such as region edges and lines, and to smooth along the image edges, removing gaps caused by noise. In its example an 8-bit grayscale Rhodamine image goes in and an 8-bit grayscale image comes out.

Two settings show how the engine measures structure. It smooths the image with a Gaussian before calculating the image Hessian, then smooths the Hessian itself, and both standard deviations default to 1. The Hessian is the table of second derivatives at each pixel. In image processing generally, its eigenvectors point along and across a line or edge, which is how a filter can tell "along" from "across".

Now the textbook register, labeled as such. The best-known anisotropic diffusion is Perona and Malik's, from 1990: ∂I/∂t = div[g(|∇I|) ∇I], where the conductance g falls from 1 toward 0 as the gradient grows, so the flow stalls at edges. That scalar form slows smoothing at an edge but can't steer it along one. Tensor forms, such as Weickert's coherence-enhancing diffusion, replace g with a matrix built from local structure, so intensity flows along lines and hardly across them.

The reference does not say which family StrataQuest's filter belongs to, and never names a method. Its settings (a Hessian, a Rotation Invariant scheme, and amplitudes named C, m and alpha, letters that also appear in Weickert's diffusivity) look like the tensor family, but that is an inference from names, not a documented fact. What the reference does document is the behavior: preserve edges and lines, smooth along edges, close gaps.

An older Anisotropic filter (v1) remains under Legacy, kept for compatibility with older projects. It has the same first five settings and no diffusion amplitudes.

Source: StrataQuest 8.1 Engines Reference, ch. 6.3 (Anisotropic filter pp. 113–115; Bilateral filter pp. 119–120; Gauss filter p. 129; Median filter p. 135; Membrane detector p. 138) and ch. 6.12 (p. 351).

The Bilateral filterLoading... also smooths while keeping edges, but in one pass. Each output pixel is a weighted average of its neighbors, weighted by distance and by intensity difference, inside a window set by Kernel height and Kernel width. Nothing reaches farther than that window. Diffusion generally works step by step instead, handing intensity from pixel to neighbor, so over the Diffusion time its effect can travel beyond any single neighborhood.

The deeper difference is direction. The bilateral weights depend on how far away and how different a neighbor is, not on which way it lies, so the Bilateral filter can keep an edge but has no notion of running along it. The anisotropic engine, per the reference, smooths along edges and removes gaps caused by noise. That makes it the one to try on line-like structures, such as the membranes the Membrane detectorLoading... detects.

The Median filterLoading... is the third edge-keeper in Filters: it conserves certain forms of borders and creates no new gray levels. Because the anisotropic smoothing depends on the image itself, it is not a fixed convolutionLoading..., and in image processing generally a step-by-step diffusion costs more computation than a single pass. The reference gives no timings, so compare on your own images. When the result feeds a thresholdLoading..., judge the smoothing by the mask it produces.