SourceExtraction.jl

SourceExtraction.jl provides fast, allocation-free, threshold-aware source detection and ROI extraction for microscopy and astronomical images, with CUDA support.

Quick start

using SourceExtraction

xs, ys = find_sources(image)

For fainter sources, lower the detection threshold:

xs, ys = find_sources(
    image;
    threshold_fraction = 0.05,
)

or set the threshold explicitly:

xs, ys = find_sources(
    image;
    threshold = 0.1f0,
)

Microscopy example

Detected sources are marked with rings in this simulated bead image:

See examples/microscopy_source_detection.jl for the simulation and detection code, or download the original TIFF image.

Prepared API

For repeated processing, prepare the extractor once and reuse its buffers:

extractor = SourceExtractor(
    image;
    dog = DoG(1f0, 3f0),
    detector = LocalMaximaDetector(0.1f0; radius=1),
    roi_size = (11, 11),
    max_sources = 100_000,
)

result = extract_sources!(extractor, image)

indices = source_indices(result)
scores  = source_scores(result)
rois    = source_rois(result)

The same API works with CuArrays when CUDA.jl is loaded.

Stacks with layout (x, y, frame) are processed as independent 2-D frames.

Performance

Representative benchmarks on a 2048×2048 Float32 image.

Strict local-maxima detection

SourceExtraction.jl and ImageFiltering.findlocalmaxima were verified to return the same source positions before timing.

Workload	SourceExtraction.jl	ImageFiltering	Speedup	SourceExtraction allocations
Raw strict 3×3 local maxima	28.03 ms	42.33 ms	1.51×	0 B
Threshold = 0.999	1.86 ms	43.77 ms	23.53×	0 B
Threshold = 0.990	2.42 ms	43.54 ms	18.01×	0 B
Threshold = 0.900	6.43 ms	44.26 ms	6.88×	0 B
DoG-filtered image, threshold = 0.1	5.54 ms	37.19 ms	6.71×	0 B

The larger speedups at higher thresholds come from rejecting sub-threshold pixels before neighborhood comparisons.

Astronomy-oriented comparison

On the same 2048×2048 DoG-filtered image and the same numerical threshold:

Method	Time	Sources
SourceExtraction detect only	2.11 ms	1474
SEP minimal extraction	19.70 ms	1461

These methods do not use identical source definitions: SourceExtraction.jl detects strict local maxima, while SEP performs thresholded connected-component extraction and measurements. The comparison is therefore practical rather than algorithmically exact.

For broader extraction pipelines on the same synthetic image:

Method	Time	Sources
SourceExtraction DoG + detect	15.98 ms	1474
SEP background + extract	133.99 ms	1292
SEP default extract	215.77 ms	1561
SExtractor CLI (WSL native filesystem)	116.61 ms	1291

The SEP and SExtractor rows perform additional work and should be treated as contextual pipeline comparisons rather than exact detector benchmarks.

CUDA

For 512×512 Float32 frames with DoG filtering, detection, and 11×11 ROI extraction:

Frames	DoG filtering	Detection	ROI extraction	Full pipeline
1	56.4 μs	18.1 μs	34.8 μs	96.1 μs
10	462.9 μs	85.5 μs	271.2 μs	813.3 μs
100	4.266 ms	764.9 μs	2.593 ms	7.574 ms

The 100-frame case processes about 510,000 detected sources and extracts 11×11 ROIs in about 7.6 ms.

See benchmark/ for the complete benchmark setup.

License

MIT