phasecongruency 1.0.4

Phase based feature detection and image enhancement

phasecongruency

Python package for phase-congruency-based feature detection and phase-preserving image enhancement.

Installation

From source with pip:

pip install .

Editable install for development:

pip install -e ".[dev]"

Or with uv:

uv sync --extra dev

GPU backend (PyTorch / Apple Metal MPS):

pip install "phasecongruency[gpu]"

Quick Start

from phasecongruency import phasecongmono, phasecong3, phasesym, step2line

img = step2line(256)

# Fast monogenic phase congruency (defaults)
PC, orientation, phase_type, T = phasecongmono(img)

# Tuned monogenic phase congruency (all key parameters configurable)
PC2, orientation2, phase_type2, T2 = phasecongmono(
    img,
    nscale=5,
    minwavelength=3,
    mult=2.1,
    sigmaonf=0.55,
    k=3.0,
    noisemethod=-1,
    cutoff=0.5,
    g=10.0,
    deviationgain=1.5,
)

# Oriented phase congruency (edge/corner moments)
M, m, or_, feat_type, EO, T3 = phasecong3(
    img,
    nscale=4,
    norient=6,
    minwavelength=3,
    mult=2.1,
    sigmaonf=0.55,
    k=2.0,
    cutoff=0.5,
    g=10.0,
    noisemethod=-1,
)

# Phase symmetry (blob/line-like structures)
phSym, orient, totalEnergy, Tps = phasesym(img)

# Optional GPU backend (auto-selects MPS on Apple Silicon if available)
PC_gpu, orientation_gpu, phase_type_gpu, T_gpu = phasecongmono(
    img,
    backend="auto",
)

# Force Apple Metal backend explicitly
M_gpu, m_gpu, or_gpu, feat_gpu, EO_gpu, T3_gpu = phasecong3(
    img,
    backend="torch-mps",
)

backend="auto" falls back to NumPy if PyTorch is not installed.

When PyTorch is available, backend="auto" benchmarks NumPy vs torch once per function + image shape + parameter set, then caches the winner on disk for reuse across runs.

• Default cache path:
  • macOS/Linux: ~/.cache/phasecongruency/auto_backend_cache.json
  • Windows: %LOCALAPPDATA%\\phasecongruency\\auto_backend_cache.json
• Override cache path with env var:
  • PHASECONGRUENCY_AUTO_CACHE_PATH=/path/to/cache.json

Parameter Configurability

Core algorithm parameters are exposed as keyword arguments in Python.

• Use keyword arguments on: phasecongmono, phasecong3, phasesymmono, phasesym, ppdrc, ppdenoise, monofilt, gaborconvolve, highpassmonogenic, and bandpassmonogenic.
• Default values are defined in each function signature.
• GPU-capable functions also accept backend ("numpy", "torch", "torch-mps", "auto") and optional device.
• Inspect any callable signature directly:

from inspect import signature
from phasecongruency import phasecongmono
print(signature(phasecongmono))

Run Tests

UV_CACHE_DIR=.uv-cache uv run pytest tests -q
# or
pytest tests -q

Scope

This repository now maintains the Python implementation only. The historical Julia transpilation/validation workflow has been retired from the active tree.

Validation Provenance

A Julia-vs-Python fidelity validation was completed before this cleanup. See:

• docs/validation.md

Attribution

This work is based on Peter Kovesi's phase congruency methods and was originally transpiled from:

• ImagePhaseCongruency.jl