ComfyUI_PhaseCongruencyCorner

ComfyUI Custom Node — Phase Congruency Corner

A ComfyUI custom node that detects corners using Phase Congruency, an illumination-invariant and contrast-invariant feature detector based on the coherence of Fourier phase components across multiple scales and orientations (Kovesi, 1999).

What is Phase Congruency Corner Detection?

Classical corner detectors (Harris, Shi-Tomasi, FAST) rely on intensity gradients, so their output changes with lighting conditions and local contrast. Phase Congruency Corner Detection takes a fundamentally different approach:

A corner is detected at locations where Fourier components are maximally in phase across all orientations simultaneously — regardless of amplitude.

• Edges have high phase congruency in one dominant orientation.
• Corners have high phase congruency in every orientation at once.

This makes the detector:

• Illumination invariant — absolute brightness has no effect
• Contrast invariant — works equally on high-contrast and low-contrast regions
• Perceptually meaningful — closely matches human visual corner perception
• Theoretically grounded — based on Kovesi (1999) using log-Gabor filters

Node

| Property | Value | |---|---| | Node name | Phase Congruency Corner | | Category | image/filters | | Input | IMAGE (RGB or grayscale, any resolution) | | Output | IMAGE (grayscale corner map, white = strong corners) |

Parameters

| Parameter | Type | Default | Range | Description | |-----------|------|---------|-------|-------------| | image | IMAGE | — | — | Input image tensor | | nscale | INT | 4 | 2–8 | Number of log-Gabor filter scales | | norient | INT | 6 | 2–8 | Number of filter orientations | | min_wavelength | INT | 3 | 2–20 | Minimum wavelength of finest-scale filter (pixels) | | mult | FLOAT | 2.1 | 1.5–4.0 | Multiplicative factor between successive filter scales | | sigma_on_f | FLOAT | 0.55 | 0.1–1.0 | Bandwidth of each log-Gabor filter (ratio σ/f₀) | | k | FLOAT | 2.0 | 0.5–10.0 | Noise threshold in std devs above mean (Rayleigh) | | cutoff | FLOAT | 0.5 | 0.1–0.9 | Butterworth low-pass filter cutoff frequency |

Algorithm

Step 1 — Per-orientation Phase Congruency maps

For each of the norient orientations θ_i:

1. Build a log-Gabor bandpass filter at each of the nscale scales, multiplied by an angular spread function (raised cosine).
2. Apply each filter in the frequency domain (FFT) to obtain complex responses.
3. Sum complex responses across all scales → energy vector (E_real, E_imag).
4. Estimate noise threshold T via Rayleigh statistics on finest-scale amplitude: T = mean_noise + k × noise_std
5. Compute per-orientation phase congruency: PC_i = max(|energy| − T, 0) / (Σ amplitude + ε)

Step 2 — Structure matrix

Accumulate a symmetric 2×2 structure matrix at each pixel:

M = Σ_i  PC_i · [[cos²θ_i,    cosθ_i·sinθ_i],
                  [cosθ_i·sinθ_i,    sin²θ_i ]]

This is analogous to the Harris structure matrix, but uses phase congruency values instead of intensity gradients.

Step 3 — Corner response

The corner strength is the minimum eigenvalue of M:

λ_min = (Mxx + Myy) / 2 − sqrt(((Mxx − Myy) / 2)² + Mxy²)

A large λ_min means high phase congruency in every spatial direction, which is the geometric signature of a corner point.

Comparison: Edge vs. Corner

| | Phase Congruency Edge | Phase Congruency Corner | |---|---|---| | Strong response when | PC is high in one direction | PC is high in all directions | | Aggregation method | Mean of per-orientation PC maps | Min. eigenvalue of structure matrix | | Typical output | Thin edge contours | Isolated corner blobs |

Comparison with Other Corner Detectors

| Method | Illumination Invariant | Contrast Invariant | Multi-scale | |--------|----------------------|-------------------|-------------| | Harris | No | No | No | | Shi-Tomasi | No | No | No | | FAST | No | No | No | | Phase Congruency Corner | Yes | Yes | Yes |

Parameter Tuning Guide

| Goal | Adjustment | |------|-----------| | Detect finer / more localised corners | Decrease min_wavelength | | Detect coarser / broader junctions | Increase min_wavelength | | More sensitive (at risk of noise) | Decrease k (e.g. 1.0) | | Cleaner output, fewer false corners | Increase k (e.g. 4.0–6.0) | | Better isotropy across all angles | Increase norient (e.g. 8) | | Richer multi-scale information | Increase nscale (e.g. 6) | | Reduce high-frequency ringing | Decrease cutoff (e.g. 0.4) |

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/bemoregt/ComfyUI_PCPoint.git

Restart ComfyUI. The node will appear under image/filters → Phase Congruency Corner.

Requirements

All dependencies are already present in a standard ComfyUI installation:

numpy
torch

Reference

Kovesi, P. (1999). Image Features from Phase Congruency. Videre: Journal of Computer Vision Research, 1(3).

License

MIT License — free for personal and commercial use.