DM-12943: Add utilities for detection when noise is correlated.

include/lsst/meas/algorithms/CorrelatedNoiseDetection.h

@@ -0,0 +1,121 @@
+// -*- lsst-c++ -*-
+/*
+ * LSST Data Management System
+ * Copyright 2017 LSST/AURA.
+ *
+ * This product includes software developed by the
+ * LSST Project (http://www.lsst.org/).
+ *
+ * This program is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the LSST License Statement and
+ * the GNU General Public License along with this program.  If not,
+ * see <http://www.lsstcorp.org/LegalNotices/>.
+ */
+
+#ifndef LSST_MEAS_ALGORITHMS_CorrelatedNoiseDetection_h_INCLUDED
+#define LSST_MEAS_ALGORITHMS_CorrelatedNoiseDetection_h_INCLUDED
+
+#include <string>
+#include <vector>
+
+#include "lsst/afw/image/MaskedImage.h"
+
+namespace lsst { namespace meas { namespace algorithms {
+
+/**
+ * Estimate the noise correlation kernel of an image, assuming it is stationary.
+ *
+ * @param[in]  image          The image to be measured.  The empirical covariance
+ *                            will be divided by the values in image's variance plane,
+ *                            and its mask will be used to reject pixels containing
+ *                            objects or artifacts.
+ * @param[in]  radius         Distance in pixels to which the correlation is measured
+ *                            on either side; the returned image will have dimensions
+ *                            (2*radius + 1, 2*radius + 1).
+ * @param[in]  badBitMask     A bit mask indicating pixels that should not be included
+ *                            in the measurement.  Should generally include at least
+ *                            DETECTED.
+ *
+ * @return  the noise correlation kernel: an image in which the central pixel
+ *          represents the fraction of the total variance/covariance in the variance,
+ *          and neighboring pixels contain the correlation at different offsets.
+ */
+afw::image::Image<float> measureCorrelationKernel(
+    afw::image::MaskedImage<float> const & image,
+    int radius,
+    afw::image::MaskPixel badBitMask
+);
+
+/**
+ * Estimate the noise correlation kernel of an image, assuming it is stationary.
+ *
+ * @param[in]  image          The image to be measured.  The empirical covariance
+ *                            will be divided by the values in image's variance plane,
+ *                            and its mask will be used to reject pixels containing
+ *                            objects or artifacts.
+ * @param[in]  radius         Distance in pixels to which the correlation is measured
+ *                            on either side; the returned image will have dimensions
+ *                            (2*radius + 1, 2*radius + 1).
+ * @param[in]  badMaskPlanes  A list of mask planenames indicating pixels that
+ *                            should not be included in the measurement.
+ *                            Should generally include at least DETECTED.
+ *
+ * @return  the noise correlation kernel: an image in which the central pixel
+ *          represents the fraction of the total variance/covariance in the variance,
+ *          and neighboring pixels contain the correlation at different offsets.
+ */
+afw::image::Image<float> measureCorrelationKernel(
+    afw::image::MaskedImage<float> const & image,
+    int radius,
+    std::vector<std::string> const & badMaskPlanes
+);
+
+
+/**
+ * Fit an optimal detection kernel for a PSF that corrects for correlated noise.
+ *
+ * This is the same as the kernel that yields the PSF when convolved with the
+ * correlation kernel.
+ *
+ * The returned kernel cannot be used in detection in quite the same way as the
+ * PSF is used in detection on images with noise correlated noise.
+ * In the uncorrelated noise case with image @f$z@f$, PSF @f$\phi@f$, and
+ * per-pixel variance @f$\sigma^2@f$, the significance image is
+ * @f[
+ *    \nu = \frac{\phi \ast z}{\sigma \sqrt{\phi \cdot \phi}}
+ * @f]
+ * In the correlated noise case, with @f$\psi@f$ the kernel computed by this
+ * function, the significance image is
+ * @f[
+ *    \nu = \frac{\psi \ast z}{\sigma \sqrt{\phi \cdot \psi}}
+ * @f]
+ * (note the difference in the denominator).
+ *
+ * @param[in]  psf          Image of the PSF model.
+ * @param[in]  correlation  Noise correlation kernel, of the sort estimated by
+ *                          measureCorrelationKernel.
+ * @param[in]  radius       Radius of the detection kernel image in pixels.
+ *                          The returned image will have dimensions
+ *                          (2*radius + 1, 2*radius + 1).
+ *
+ * @return  an image containing the optimal detection kernel.
+ */
+afw::image::Image<double> fitGeneralDetectionKernel(
+    afw::image::Image<double> const & psf,
+    afw::image::Image<float> const & correlation,
+    int radius
+);
+
+
+}}} // namespace lsst::meas::algorithms
+
+#endif // !LSST_MEAS_ALGORITHMS_CorrelatedNoiseDetection_h_INCLUDED

python/lsst/meas/algorithms/SConscript

@@ -4,6 +4,7 @@ scripts.BasicSConscript.pybind11(["binnedWcs",
                                   "cr",
                                   "coaddBoundedField",
                                   "coaddPsf/coaddPsf",
+                                  "correlatedNoiseDetection",
                                   "doubleGaussianPsf",
                                   "imagePsf",
                                   "interp",

python/lsst/meas/algorithms/__init__.py

@@ -63,6 +63,7 @@
 from .loadIndexedReferenceObjects import *
 from .indexerRegistry import *
 from .reserveSourcesTask import *
+from .correlatedNoiseDetection import *
 
 from .version import *
 

python/lsst/meas/algorithms/correlatedNoiseDetection.cc

@@ -0,0 +1,66 @@
+/*
+ * LSST Data Management System
+ * Copyright 2017 LSST/AURA.
+ *
+ * This product includes software developed by the
+ * LSST Project (http://www.lsst.org/).
+ *
+ * This program is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the LSST License Statement and
+ * the GNU General Public License along with this program.  If not,
+ * see <https://www.lsstcorp.org/LegalNotices/>.
+ */
+#include "pybind11/pybind11.h"
+#include "pybind11/stl.h"
+
+#include "lsst/meas/algorithms/CorrelatedNoiseDetection.h"
+
+namespace py = pybind11;
+using namespace pybind11::literals;
+
+namespace lsst {
+namespace meas {
+namespace algorithms {
+
+PYBIND11_PLUGIN(correlatedNoiseDetection) {
+    py::module::import("lsst.afw.image");
+
+    py::module mod("correlatedNoiseDetection");
+
+    mod.def(
+        "measureCorrelationKernel",
+        (afw::image::Image<float> (*)(
+            afw::image::MaskedImage<float> const &,
+            int,
+            afw::image::MaskPixel
+        ))&measureCorrelationKernel,
+        "image"_a, "radius"_a, "badBitMask"_a
+    );
+
+    mod.def(
+        "measureCorrelationKernel",
+        (afw::image::Image<float> (*)(
+            afw::image::MaskedImage<float> const &,
+            int,
+            std::vector<std::string> const &
+        ))&measureCorrelationKernel,
+        "image"_a, "radius"_a, "badMaskPlanes"_a
+    );
+
+    mod.def("fitGeneralDetectionKernel", &fitGeneralDetectionKernel, "psf"_a, "correlations"_a, "radius"_a);
+
+    return mod.ptr();
+}
+
+}  // algorithms
+}  // meas
+}  // lsst

src/CorrelatedNoiseDetection.cc

@@ -0,0 +1,192 @@
+// -*- LSST-C++ -*-
+/*
+ * LSST Data Management System
+ * Copyright 2017 LSST/AURA.
+ *
+ * This product includes software developed by the
+ * LSST Project (http://www.lsst.org/).
+ *
+ * This program is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the LSST License Statement and
+ * the GNU General Public License along with this program.  If not,
+ * see <http://www.lsstcorp.org/LegalNotices/>.
+ */
+
+#include "lsst/pex/exceptions.h"
+#include "lsst/afw/math/LeastSquares.h"
+#include "lsst/meas/algorithms/CorrelatedNoiseDetection.h"
+
+namespace lsst { namespace meas { namespace algorithms {
+
+
+afw::image::Image<float> measureCorrelationKernel(
+    afw::image::MaskedImage<float> const & mi,
+    int radius,
+    afw::image::MaskPixel badBitMask
+) {
+    afw::image::Image<float> result(2*radius + 1, 2*radius + 1);
+    afw::image::Image<int> count(result.getDimensions());
+    int const width = mi.getWidth();
+    int const height = mi.getHeight();
+    // iterate over pixels in the MaskedImage, skipping any that meet our bad mask criteria
+    for (int y1 = 0; y1 < height; ++y1) {
+        int const y2a = std::max(0, y1 - radius);
+        int const y2b = std::min(height, y1 + radius + 1);
+        for (int x1 = 0; x1 < width; ++x1) {
+            if ((*mi.getMask())(x1, y1) & badBitMask) {
+                continue;
+            }
+            float z = (*mi.getImage())(x1, y1) / (*mi.getVariance())(x1, y1);
+            // iterate over neighboring pixels, with bounds set to avoid image boundaries
+            int const x2a = std::max(0, x1 - radius);
+            int const x2b = std::min(height, x1 + radius + 1);
+            for (int y2 = y2a; y2 < y2b; ++y2) {
+                auto miIter = mi.row_begin(y2) + x2a;
+                auto outIter = result.row_begin(radius + y2 - y1) + radius + x2a - x1;
+                auto countIter = count.row_begin(radius + y2 - y1) + radius + x2a - x1;
+                for (int x2 = x2a; x2 < x2b; ++x2, ++miIter, ++outIter, ++countIter) {
+                    if (miIter.mask() & badBitMask) {
+                        continue;
+                    }
+                    *outIter += z*miIter.image();
+                    *countIter += 1;
+                }
+            }
+        }
+    }
+    result.getArray().deep() /= count.getArray();
+    result.setXY0(-radius, -radius);
+    return result;
+}
+
+afw::image::Image<float> measureCorrelationKernel(
+    afw::image::MaskedImage<float> const & image,
+    int radius,
+    std::vector<std::string> const & badMaskPlanes
+) {
+    afw::image::MaskPixel mask = 0x0;
+    for (auto plane : badMaskPlanes) {
+        mask |= image.getMask()->getPlaneBitMask(plane);
+    }
+    return measureCorrelationKernel(image, radius, mask);
+}
+
+
+namespace {
+
+afw::geom::Extent2I getOddBoxHalfWidth(afw::geom::Box2I const & box, std::string const & name) {
+    afw::geom::Extent2I r((box.getWidth() - 1) / 2, (box.getHeight() - 1) / 2);
+    if (r.getX()*2 + 1 != box.getWidth()) {
+        throw LSST_EXCEPT(
+            pex::exceptions::InvalidParameterError,
+            name + " image width must be an odd integer"
+        );
+    }
+    if (r.getY()*2 + 1 != box.getHeight()) {
+        throw LSST_EXCEPT(
+            pex::exceptions::InvalidParameterError,
+            name + " image height must be an odd integer"
+        );
+    }
+    return r;
+}
+
+} // anonymous
+
+
+afw::image::Image<double> fitGeneralDetectionKernel(
+    afw::image::Image<double> const & psf,
+    afw::image::Image<float> const & correlation,
+    int radius
+) {
+    auto const psfR = getOddBoxHalfWidth(psf.getBBox(), "PSF");
+    auto const corrR = getOddBoxHalfWidth(correlation.getBBox(), "Correlation kernel");
+    afw::geom::Extent2I const outR(radius, radius);
+    if (psfR.getX() <= corrR.getX()) {
+        throw LSST_EXCEPT(
+            pex::exceptions::InvalidParameterError,
+            "PSF image width must be greater than correlation kernel width"
+        );
+    }
+    if (psfR.getY() <= corrR.getY()) {
+        throw LSST_EXCEPT(
+            pex::exceptions::InvalidParameterError,
+            "PSF image height must be greater than correlation kernel height"
+        );
+    }
+    if (psfR.getX() <= outR.getX()) {
+        throw LSST_EXCEPT(
+            pex::exceptions::InvalidParameterError,
+            "PSF image width must be greater than output kernel width"
+        );
+    }
+    if (psfR.getY() <= outR.getY()) {
+        throw LSST_EXCEPT(
+            pex::exceptions::InvalidParameterError,
+            "PSF image height must be greater than output kernel height"
+        );
+    }
+
+    int const psfN = (psfR.getX()*2 + 1)*(psfR.getY()*2 + 1);
+    int const outN = (outR.getX()*2 + 1)*(outR.getY()*2 + 1);
+
+    // Get locators at the center of each image, so we can use indices with
+    // the origin at the center and make the code easier to read.
+    auto psfL = psf.xy_at(psfR.getX(), psfR.getY());
+    auto corrL = correlation.xy_at(corrR.getX(), corrR.getY());
+
+    // rhs is the PSF image, with rows and columns flattened into one dimension
+    Eigen::VectorXd rhs = Eigen::VectorXd::Zero(psfN);
+
+    // matrix represents convolution with the correlated noise kernel image
+    Eigen::MatrixXd matrix = Eigen::MatrixXd::Zero(psfN, outN);
+    int xyN = 0;
+    for (int y = -psfR.getY(); y <= psfR.getY(); ++y) {
+        for (int x = -psfR.getX(); x <= psfR.getX(); ++x) {
+            int ijN = 0;
+            for (int i = -outR.getY(); i <= outR.getY(); ++i) {
+                for (int j = -outR.getX(); j <= outR.getX(); ++j) {
+                    // Could move these checks into the inner loop bounds for performance,
+                    // but this is easier to read and probably fast enough.
+                    if (std::abs(y - i) <= corrR.getY() && std::abs(x - j) <= corrR.getX()) {
+                        matrix(xyN, ijN) = corrL(x - j, y - i);
+                    }
+                    ++ijN;
+                }
+            }
+            rhs[xyN] = psfL(x, y);
+            ++xyN;
+        }
+    }
+
+    // solve for the kernel that produces the PSF when convolved with the
+    // noise correlation kernel
+    auto lstsq = afw::math::LeastSquares::fromDesignMatrix(matrix, rhs);
+    auto solution = lstsq.getSolution();
+
+    // copy the result from the flattened solution vector into an image
+    afw::image::Image<double> result(outR.getX()*2 + 1, outR.getY()*2 + 1);
+    auto outL = result.xy_at(outR.getX(), outR.getY());
+    xyN = 0;
+    for (int y = -outR.getY(); y <= outR.getY(); ++y) {
+        for (int x = -outR.getX(); x <= outR.getX(); ++x) {
+            outL(x, y) = solution[xyN];
+            ++xyN;
+        }
+    }
+
+    result.setXY0(-outR.getX(), -outR.getY());
+    return result;
+}
+
+
+}}} // namespace lsst::meas::algorithms