// SPDX-FileCopyrightText: 2026 Filip Leonarski, Paul Scherrer Institute // SPDX-License-Identifier: GPL-3.0-only #include #include #include "../image_analysis/spot_finding/AdaptiveThreshold.h" using namespace adaptive_threshold; namespace { // Poisson upper tail P(X >= k) for mean mu, summed directly - an independent reference for the // threshold's defining property. double PoissonUpperTail(double mu, int k) { if (k <= 0) return 1.0; double pmf = std::exp(-mu); double cdf = pmf; for (int i = 1; i < k; i++) { pmf *= mu / i; cdf += pmf; } return std::max(0.0, 1.0 - cdf); } } TEST_CASE("AdaptiveThreshold_NormalQuantile", "[SpotFinding]") { // Textbook values of the inverse standard-normal CDF. CHECK(NormalQuantile(0.5) == Catch::Approx(0.0).margin(1e-9)); CHECK(NormalQuantile(0.975) == Catch::Approx(1.959964).margin(1e-5)); CHECK(NormalQuantile(0.99) == Catch::Approx(2.326348).margin(1e-5)); CHECK(NormalQuantile(1.0 - 1e-6) == Catch::Approx(4.753424).margin(1e-4)); // Symmetric about 0.5, and monotonically increasing. for (const double p: {1e-8, 1e-4, 0.01, 0.2, 0.45}) CHECK(NormalQuantile(1.0 - p) == Catch::Approx(-NormalQuantile(p)).margin(1e-6)); CHECK(NormalQuantile(0.6) > NormalQuantile(0.55)); CHECK(NormalQuantile(1e-3) < NormalQuantile(1e-2)); // Degenerate arguments stay finite: the finders divide a tolerated-false-pixel count by the pixel // count, so p can legitimately arrive at the very edge of (0, 1). CHECK(std::isfinite(NormalQuantile(0.0))); CHECK(std::isfinite(NormalQuantile(1.0))); CHECK(NormalQuantile(0.0) < 0.0); CHECK(NormalQuantile(1.0) > 0.0); } TEST_CASE("AdaptiveThreshold_PoissonThreshold", "[SpotFinding]") { const double p = 1e-5; const float z = static_cast(NormalQuantile(1.0 - p)); // The defining property: the returned count is the SMALLEST whose upper tail is within p. for (const double mu: {1e-6, 0.1, 1.0, 3.0, 10.0, 40.0}) { const int thr = static_cast(PoissonThreshold(mu, p, z)); CHECK(PoissonUpperTail(mu, thr) <= p); CHECK(PoissonUpperTail(mu, thr - 1) > p); } // Non-decreasing in the background level. float prev = 0.0f; for (const double mu: {1e-6, 0.01, 0.1, 0.5, 1.0, 2.0, 5.0, 20.0, 45.0}) { const float thr = PoissonThreshold(mu, p, z); CHECK(thr >= prev); prev = thr; } // Above mu = 50 it short-circuits to the Gaussian form mu + z sqrt(mu). CHECK(PoissonThreshold(100.0, p, z) == Catch::Approx(100.0 + z * 10.0).epsilon(1e-5)); // A tighter operating point (smaller p) can only raise the threshold. CHECK(PoissonThreshold(5.0, 1e-8, static_cast(NormalQuantile(1.0 - 1e-8))) >= PoissonThreshold(5.0, 1e-2, static_cast(NormalQuantile(1.0 - 1e-2)))); } TEST_CASE("AdaptiveThreshold_RingThreshold", "[SpotFinding]") { const double p = 1e-5; const float z = static_cast(NormalQuantile(1.0 - p)); // Never below the read-noise-aware Gaussian arm, which is what keeps an empty ring's threshold // off zero - a per-ring sigma alone would collapse there and flood the frame with noise spots. for (const float mean: {0.0f, 0.5f, 5.0f, 50.0f}) { for (const float sigma: {0.0f, 1.0f, 7.0f}) { const float gauss = mean + z * std::sqrt(sigma * sigma + READ * READ); CHECK(RingThreshold(mean, sigma, p, z) >= Catch::Approx(gauss).epsilon(1e-6)); } } CHECK(RingThreshold(0.0f, 0.0f, p, z) >= z * READ); // Non-decreasing in the background mean and in the background scatter. CHECK(RingThreshold(20.0f, 4.0f, p, z) > RingThreshold(2.0f, 4.0f, p, z)); CHECK(RingThreshold(5.0f, 9.0f, p, z) > RingThreshold(5.0f, 1.0f, p, z)); // Where the background is countable and quiet, Poisson significance is the binding arm: a ring // with mean 1 and no measured scatter must still demand several photons. CHECK(RingThreshold(1.0f, 0.0f, p, z) > 1.0f + z * READ); // A ring whose scatter is far above Poisson (flat-field / read excess) is set by the Gaussian arm. CHECK(RingThreshold(10.0f, 30.0f, p, z) == Catch::Approx(10.0f + z * std::sqrt(900.0f + READ * READ)).epsilon(1e-6)); }