GitHub

SourceExtraction.jl

SourceExtraction.jl provides fast source detection and region-of-interest (ROI) extraction for microscopy and astronomical images. It combines Difference-of-Gaussians (DoG) filtering with thresholded local-maximum detection, supports 2-D images and (row, column, frame) stacks, and can reuse its working buffers in repeated workloads.

Installation

Install the package directly from GitHub:

using Pkg
Pkg.add(url="https://github.com/hzarei4/SourceExtraction.jl")

CUDA support is provided by a package extension and is activated automatically when CUDA.jl is loaded.

Quick start

For a single 2-D image, find_sources returns separate vectors of horizontal (x, or column) and vertical (y, or row) coordinates:

using SourceExtraction

image = zeros(Float32, 64, 64)
image[20, 30] = 10
image[45, 42] = 8

xs, ys = find_sources(image; threshold=0.1f0)
sort!(collect(zip(xs, ys)))
2-element Vector{Tuple{Int64, Int64}}:
 (30, 20)
 (42, 45)

The default threshold is 25% of the strongest DoG-filtered response. Set threshold_fraction to another fraction or pass an absolute threshold as in the example above.

For a stream of same-sized images, prepare a SourceExtractor once and reuse it:

extractor = SourceExtractor(
    image;
    dog=DoG(1f0, 3f0),
    detector=LocalMaximaDetector(0.1f0),
    roi_size=(11, 11),
    max_sources=100,
)

result = extract_sources!(extractor, image)
(
    count=source_count(result),
    indices=collect(source_indices(result)),
    roi_size=size(source_rois(result)),
)
(count = 2, indices = CartesianIndex{2}[CartesianIndex(20, 30), CartesianIndex(45, 42)], roi_size = (11, 11, 2))

Results from the prepared API reference reusable storage owned by the extractor. Copy a result view if its contents must survive the next call.

See the user guide for thresholds, batched data, ROIs, CUDA, and the low-level allocation-free interface. The complete public interface is listed in the API reference.

Package features

  • One-shot and prepared APIs for different workload sizes.
  • Strict or non-strict local maxima with configurable radius and threshold.
  • Optional fixed-size ROI extraction from the original input.
  • Independent processing of every frame in 3-D image stacks.
  • Reusable CPU and CUDA buffers for performance-sensitive pipelines.