mirror of
https://github.com/opencv/opencv.git
synced 2026-11-16 06:17:53 -06:00
* imgproc: fix stackBlur results on full-range 16-bit images The 16-bit SIMD paths overflowed 32 bits in (sum * mul) >> shr and saturated when adding two 16-bit samples before the kernel multiply. Add the samples after widening, and divide as the scalar loops do, so SIMD and scalar agree and the result no longer depends on the thread count. * imgproc: scale the stackBlur thread test with the machine's thread count Restore the count with a guard so a throw can't leak it. Use an odd image width: with a width of 160, stripes at some thread counts, including 12, leave nothing to the scalar loop and the test passes on the unfixed code. * imgproc: use 64-bit sums for large stackBlur kernels on 16-bit images The largest sum is the largest absolute value times (radius + 1)^2, which no longer fits an int from radius 181 for CV_16U and radius 255 for CV_16S. Those radii now take the scalar loops with int64_t sums. Radii that fit keep the 32-bit SIMD paths. * imgproc: use 64-bit sums for large stackBlur kernels on 8-bit images
484 lines
15 KiB
C++
484 lines
15 KiB
C++
// This file is part of OpenCV project.
|
|
// It is subject to the license terms in the LICENSE file found in the top-level directory
|
|
// of this distribution and at http://opencv.org/license.html.
|
|
|
|
/*
|
|
StackBlur - a fast almost Gaussian Blur
|
|
Theory: http://underdestruction.com/2004/02/25/stackblur-2004
|
|
The code has been borrowed from (https://github.com/flozz/StackBlur).
|
|
|
|
Below is the original copyright
|
|
*/
|
|
|
|
/*
|
|
Copyright (c) 2010 Mario Klingemann
|
|
|
|
Permission is hereby granted, free of charge, to any person
|
|
obtaining a copy of this software and associated documentation
|
|
files (the "Software"), to deal in the Software without
|
|
restriction, including without limitation the rights to use,
|
|
copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
copies of the Software, and to permit persons to whom the
|
|
Software is furnished to do so, subject to the following
|
|
conditions:
|
|
|
|
The above copyright notice and this permission notice shall be
|
|
included in all copies or substantial portions of the Software.
|
|
|
|
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
|
|
EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
|
|
OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
|
|
NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
|
|
HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
|
|
WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
|
FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
|
|
OTHER DEALINGS IN THE SOFTWARE.
|
|
*/
|
|
|
|
|
|
#include "test_precomp.hpp"
|
|
|
|
namespace opencv_test { namespace {
|
|
|
|
template<typename T>
|
|
void _stackblurRef(const Mat& src, Mat& dst, Size ksize)
|
|
{
|
|
CV_Assert(!src.empty());
|
|
CV_Assert(ksize.width > 0 && ksize.height > 0 && ksize.height % 2 == 1 && ksize.width % 2 == 1);
|
|
|
|
dst.create(src.size(), src.type());
|
|
const int CN = src.channels();
|
|
|
|
int rowsImg = src.rows;
|
|
int colsImg = src.cols;
|
|
int wm = colsImg - 1;
|
|
|
|
int radiusW = ksize.width / 2;
|
|
int stackLenW = ksize.width;
|
|
const float mulW = 1.0f / (((float )radiusW + 1.0f) * ((float )radiusW + 1.0f));
|
|
|
|
// Horizontal direction
|
|
std::vector<T> stack(stackLenW * CN);
|
|
for (int row = 0; row < rowsImg; row++)
|
|
{
|
|
std::vector<float> sum(CN, 0);
|
|
std::vector<float> sumIn(CN, 0);
|
|
std::vector<float> sumOut(CN, 0);
|
|
|
|
const T* srcPtr = src.ptr<T>(row);
|
|
|
|
for (int i = 0; i <= radiusW; i++)
|
|
{
|
|
for (int ci = 0; ci < CN; ci++)
|
|
{
|
|
T tmp = *(srcPtr + ci);
|
|
stack[i * CN + ci] = tmp;
|
|
sum[ci] += tmp * (i + 1);
|
|
sumOut[ci] += tmp;
|
|
}
|
|
}
|
|
|
|
for (int i = 1; i <= radiusW; i++)
|
|
{
|
|
if (i <= wm) srcPtr += CN;
|
|
for(int ci = 0; ci < CN; ci++)
|
|
{
|
|
T tmp = *(srcPtr + ci);
|
|
stack[(i + radiusW) * CN + ci] = tmp;
|
|
sum[ci] += tmp * (radiusW + 1 - i);
|
|
sumIn[ci] += tmp;
|
|
}
|
|
}
|
|
|
|
int sp = radiusW;
|
|
int xp = radiusW ;
|
|
if (xp > wm) xp = wm;
|
|
|
|
T* dstPtr = dst.ptr<T>(row);
|
|
srcPtr = src.ptr<T>(row) + xp * CN;
|
|
|
|
int stackStart= 0;
|
|
|
|
for (int i = 0; i < colsImg; i++)
|
|
{
|
|
stackStart = sp + stackLenW - radiusW;
|
|
|
|
if (stackStart >= stackLenW) stackStart -= stackLenW;
|
|
|
|
for(int ci = 0; ci < CN; ci++)
|
|
{
|
|
*(dstPtr + ci) = cv::saturate_cast<T>(sum[ci] * mulW);
|
|
sum[ci] -= sumOut[ci];
|
|
sumOut[ci] -= stack[stackStart*CN + ci];
|
|
}
|
|
|
|
const T* srcNew = srcPtr;
|
|
|
|
if(xp < wm)
|
|
srcNew += CN;
|
|
|
|
for (int ci = 0; ci < CN; ci++)
|
|
{
|
|
stack[stackStart * CN + ci] = *(srcNew + ci);
|
|
sumIn[ci] += *(srcNew + ci);
|
|
sum[ci] += sumIn[ci];
|
|
}
|
|
|
|
int sp1 = sp + 1;
|
|
if (sp1 >= stackLenW)
|
|
sp1 = 0;
|
|
|
|
for(int ci = 0; ci < CN; ci++)
|
|
{
|
|
T tmp = stack[sp1*CN + ci];
|
|
sumOut[ci] += tmp;
|
|
sumIn[ci] -= tmp;
|
|
}
|
|
|
|
dstPtr += CN;
|
|
|
|
if (xp < wm)
|
|
{
|
|
xp++;
|
|
srcPtr += CN;
|
|
}
|
|
|
|
++sp;
|
|
if (sp >= stackLenW)
|
|
sp = 0;
|
|
}
|
|
}
|
|
|
|
// Vertical direction
|
|
int hm = rowsImg - 1;
|
|
int widthElem = colsImg * CN;
|
|
int radiusH = ksize.height / 2;
|
|
int stackLenH = ksize.height;
|
|
const float mulH = 1.0f / (((float )radiusH + 1.0f) * ((float )radiusH + 1.0f));
|
|
|
|
stack.resize(stackLenH, 0);
|
|
for (int col = 0; col < widthElem; col++)
|
|
{
|
|
const T* srcPtr =dst.ptr<T>() + col;
|
|
float sum0 = 0;
|
|
float sumIn0 = 0;
|
|
float sumOut0 = 0;
|
|
|
|
for (int i = 0; i <= radiusH; i++)
|
|
{
|
|
T tmp = (T)(*srcPtr);
|
|
stack[i] = tmp;
|
|
sum0 += tmp * (i + 1);
|
|
sumOut0 += tmp;
|
|
}
|
|
|
|
for (int i = 1; i <= radiusH; i++)
|
|
{
|
|
if (i <= hm) srcPtr += widthElem;
|
|
T tmp = (T)(*srcPtr);
|
|
stack[i + radiusH] = tmp;
|
|
sum0 += tmp * (radiusH - i + 1);
|
|
sumIn0 += tmp;
|
|
}
|
|
|
|
int sp = radiusH;
|
|
int yp = radiusH;
|
|
|
|
if (yp > hm) yp = hm;
|
|
|
|
T* dstPtr = dst.ptr<T>() + col;
|
|
srcPtr = dst.ptr<T>(yp) + col;
|
|
|
|
const T* srcNew;
|
|
|
|
int stackStart = 0;
|
|
|
|
for (int i = 0; i < rowsImg; i++)
|
|
{
|
|
stackStart = sp + stackLenH - radiusH;
|
|
if (stackStart >= stackLenH) stackStart -= stackLenH;
|
|
|
|
*(dstPtr) = saturate_cast<T>(sum0 * mulH);
|
|
sum0 -= sumOut0;
|
|
sumOut0 -= stack[stackStart];
|
|
srcNew = srcPtr;
|
|
|
|
if (yp < hm)
|
|
srcNew += widthElem;
|
|
|
|
stack[stackStart] = *(srcNew);
|
|
sumIn0 += *(srcNew);
|
|
sum0 += sumIn0;
|
|
|
|
int sp1 = sp + 1;
|
|
sp1 &= -(sp1 < stackLenH);
|
|
|
|
sumOut0 += stack[sp1];
|
|
sumIn0 -= stack[sp1];
|
|
|
|
dstPtr += widthElem;
|
|
|
|
if (yp < hm)
|
|
{
|
|
yp++;
|
|
srcPtr += widthElem;
|
|
}
|
|
|
|
++sp;
|
|
if (sp >= stackLenH) sp = 0;
|
|
}
|
|
}
|
|
}
|
|
|
|
void stackBlurRef(const Mat& img, Mat& dst, Size ksize)
|
|
{
|
|
if(img.depth() == CV_8U)
|
|
_stackblurRef<uchar>(img, dst, ksize);
|
|
else if (img.depth() == CV_16S)
|
|
_stackblurRef<short>(img, dst, ksize);
|
|
else if (img.depth() == CV_16U)
|
|
_stackblurRef<ushort>(img, dst, ksize);
|
|
else if (img.depth() == CV_32F)
|
|
_stackblurRef<float>(img, dst, ksize);
|
|
else
|
|
CV_Error(Error::StsNotImplemented,
|
|
("Unsupported Mat type in stackBlurRef, "
|
|
"the supported formats are: CV_8U, CV_16U, CV_16S and CV_32F."));
|
|
}
|
|
|
|
std::vector<Size> kernelSizeVec = {
|
|
Size(3, 3),
|
|
Size(5, 5),
|
|
Size(101, 101),
|
|
Size(3, 9)
|
|
};
|
|
|
|
typedef testing::TestWithParam<tuple<int, int, int> > StackBlur;
|
|
|
|
TEST_P (StackBlur, regression)
|
|
{
|
|
Mat img_ = imread(findDataFile("shared/fruits.png"), 1);
|
|
const int cn = get<0>(GetParam());
|
|
const int kIndex = get<1>(GetParam());
|
|
const int dtype = get<2>(GetParam());
|
|
|
|
Size ksize = kernelSizeVec[kIndex];
|
|
|
|
Mat img, dstRef, dst;
|
|
convert(img_, img, dtype);
|
|
|
|
vector<Mat> channels;
|
|
split(img, channels);
|
|
channels.push_back(channels[0]); // channels size is 4.
|
|
|
|
Mat imgCn;
|
|
if (cn == 1)
|
|
imgCn = channels[0];
|
|
else if (cn == 4)
|
|
merge(channels, imgCn);
|
|
else
|
|
imgCn = img;
|
|
|
|
stackBlurRef(imgCn, dstRef, ksize);
|
|
stackBlur(imgCn, dst, ksize);
|
|
EXPECT_LE(cvtest::norm(dstRef, dst, NORM_INF), 2.);
|
|
}
|
|
|
|
INSTANTIATE_TEST_CASE_P(Imgproc, StackBlur,
|
|
testing::Combine(
|
|
testing::Values(1, 3, 4),
|
|
testing::Values(0, 1, 2, 3),
|
|
testing::Values(CV_8U, CV_16S, CV_16U, CV_32F)
|
|
)
|
|
);
|
|
|
|
typedef testing::TestWithParam<tuple<int> > StackBlur_GaussianBlur;
|
|
|
|
// StackBlur should produce similar results as GaussianBlur output.
|
|
TEST_P(StackBlur_GaussianBlur, compare)
|
|
{
|
|
Mat img_ = imread(findDataFile("shared/fruits.png"), 1);
|
|
const int dtype = get<0>(GetParam());
|
|
|
|
Size ksize(3, 3);
|
|
Mat img, dstS, dstG;
|
|
convert(img_, img, dtype);
|
|
|
|
stackBlur(img, dstS, ksize);
|
|
GaussianBlur(img, dstG, ksize, 0);
|
|
|
|
EXPECT_LE(cvtest::norm(dstS, dstG, NORM_INF), 13.);
|
|
}
|
|
|
|
INSTANTIATE_TEST_CASE_P(Imgproc, StackBlur_GaussianBlur, testing::Values(CV_8U, CV_16S, CV_16U, CV_32F));
|
|
|
|
TEST(Imgproc_StackBlur, regression_28233)
|
|
{
|
|
Mat src1(1, 1, CV_8UC1, Scalar(123));
|
|
Mat dst1;
|
|
EXPECT_NO_THROW(stackBlur(src1, dst1, Size(9, 1)));
|
|
EXPECT_EQ(dst1.at<uchar>(0, 0), 123);
|
|
|
|
Mat src2(3, 3, CV_8UC1, Scalar(50));
|
|
Mat dst2;
|
|
EXPECT_NO_THROW(stackBlur(src2, dst2, Size(11, 11)));
|
|
EXPECT_EQ(dst2.at<uchar>(1, 1), 50);
|
|
}
|
|
|
|
typedef testing::TestWithParam<tuple<int, int, int> > StackBlur_FullRange;
|
|
|
|
// A constant image stays constant, however large its values, and full-range noise matches the reference.
|
|
TEST_P(StackBlur_FullRange, regression)
|
|
{
|
|
const int depth = get<0>(GetParam());
|
|
const int cn = get<1>(GetParam());
|
|
const int kernel = get<2>(GetParam());
|
|
const Size ksize(kernel, kernel);
|
|
const double maxVal = depth == CV_16U ? 65535 : 32767;
|
|
const double minVal = depth == CV_16U ? 0 : -32768;
|
|
|
|
for (double value : {maxVal, minVal, maxVal / 2})
|
|
{
|
|
Mat constant(64, 96, CV_MAKETYPE(depth, cn), Scalar::all(value)), dst;
|
|
stackBlur(constant, dst, ksize);
|
|
EXPECT_LE(cvtest::norm(constant, dst, NORM_INF), 1.) << "value=" << value;
|
|
}
|
|
|
|
// stackBlurRef accumulates in float, which drifts at this range, so the reference is the
|
|
// triangular kernel stackBlur stands for, computed in double.
|
|
const int radius = kernel / 2;
|
|
Mat tri(kernel, 1, CV_64F);
|
|
for (int i = 0; i < kernel; i++)
|
|
tri.at<double>(i) = (radius + 1 - std::abs(i - radius)) / double((radius + 1) * (radius + 1));
|
|
|
|
Mat noise(64, 96, CV_MAKETYPE(depth, cn)), dst, ref, dst64;
|
|
randu(noise, Scalar::all(minVal), Scalar::all(maxVal));
|
|
sepFilter2D(noise, ref, CV_64F, tri, tri, Point(-1, -1), 0, BORDER_REPLICATE);
|
|
stackBlur(noise, dst, ksize);
|
|
dst.convertTo(dst64, CV_64F);
|
|
const Rect inner(radius, radius, noise.cols - 2 * radius, noise.rows - 2 * radius);
|
|
EXPECT_LE(cvtest::norm(ref(inner), dst64(inner), NORM_INF), 2.);
|
|
}
|
|
|
|
INSTANTIATE_TEST_CASE_P(Imgproc, StackBlur_FullRange,
|
|
testing::Combine(
|
|
testing::Values(CV_16U, CV_16S),
|
|
testing::Values(1, 3, 4),
|
|
testing::Values(3, 5, 7, 15, 51)
|
|
)
|
|
);
|
|
|
|
typedef testing::TestWithParam<tuple<int, int> > StackBlur_LargeKernel;
|
|
|
|
// The sum no longer fits 32 bits from radius 181 for CV_16U and radius 255 for CV_16S.
|
|
TEST_P(StackBlur_LargeKernel, regression)
|
|
{
|
|
const int depth = get<0>(GetParam());
|
|
const int kernel = get<1>(GetParam());
|
|
const double maxVal = depth == CV_16U ? 65535 : 32767;
|
|
const double minVal = depth == CV_16U ? 0 : -32768;
|
|
const int side = kernel + 32;
|
|
|
|
for (double value : {maxVal, minVal, maxVal / 2})
|
|
{
|
|
Mat constant(side, side, CV_MAKETYPE(depth, 3), Scalar::all(value)), dst;
|
|
for (const Size& ksize : {Size(kernel, kernel), Size(kernel, 1), Size(1, kernel)})
|
|
{
|
|
stackBlur(constant, dst, ksize);
|
|
EXPECT_LE(cvtest::norm(constant, dst, NORM_INF), 1.) << "value=" << value << " ksize=" << ksize;
|
|
}
|
|
}
|
|
|
|
const int radius = kernel / 2;
|
|
Mat tri(kernel, 1, CV_64F);
|
|
for (int i = 0; i < kernel; i++)
|
|
tri.at<double>(i) = (radius + 1 - std::abs(i - radius)) / double((radius + 1) * (radius + 1));
|
|
|
|
Mat noise(side, side, depth), dst, ref, dst64;
|
|
randu(noise, Scalar::all(minVal), Scalar::all(maxVal));
|
|
sepFilter2D(noise, ref, CV_64F, tri, tri, Point(-1, -1), 0, BORDER_REPLICATE);
|
|
stackBlur(noise, dst, Size(kernel, kernel));
|
|
dst.convertTo(dst64, CV_64F);
|
|
const Rect inner(radius, radius, noise.cols - 2 * radius, noise.rows - 2 * radius);
|
|
EXPECT_LE(cvtest::norm(ref(inner), dst64(inner), NORM_INF), 2.);
|
|
}
|
|
|
|
INSTANTIATE_TEST_CASE_P(Imgproc, StackBlur_LargeKernel,
|
|
testing::Combine(
|
|
testing::Values(CV_16U, CV_16S),
|
|
testing::Values(361, 363, 509, 511, 513)
|
|
)
|
|
);
|
|
|
|
// For CV_8U the sum no longer fits 32 bits from radius 2901.
|
|
TEST(Imgproc_StackBlur, large_kernel_8u)
|
|
{
|
|
for (int kernel : {5801, 5803})
|
|
{
|
|
const int radius = kernel / 2;
|
|
for (double value : {255., 128.})
|
|
{
|
|
Mat row(3, kernel + 32, CV_8UC1, Scalar::all(value)), col(kernel + 32, 3, CV_8UC1, Scalar::all(value)), dst;
|
|
stackBlur(row, dst, Size(kernel, 1));
|
|
EXPECT_LE(cvtest::norm(row, dst, NORM_INF), 1.) << "value=" << value << " kernel=" << kernel;
|
|
stackBlur(col, dst, Size(1, kernel));
|
|
EXPECT_LE(cvtest::norm(col, dst, NORM_INF), 1.) << "value=" << value << " kernel=" << kernel;
|
|
}
|
|
|
|
Mat tri(1, kernel, CV_64F), one(1, 1, CV_64F, Scalar::all(1.));
|
|
for (int i = 0; i < kernel; i++)
|
|
tri.at<double>(i) = (radius + 1 - std::abs(i - radius)) / double((radius + 1) * (radius + 1));
|
|
|
|
Mat noise(3, 2 * kernel, CV_8UC1), dst, ref, dst64;
|
|
randu(noise, Scalar::all(0), Scalar::all(256));
|
|
sepFilter2D(noise, ref, CV_64F, tri, one, Point(-1, -1), 0, BORDER_REPLICATE);
|
|
stackBlur(noise, dst, Size(kernel, 1));
|
|
dst.convertTo(dst64, CV_64F);
|
|
const Rect inner(radius, 0, noise.cols - 2 * radius, noise.rows);
|
|
EXPECT_LE(cvtest::norm(ref(inner), dst64(inner), NORM_INF), 1.) << "kernel=" << kernel;
|
|
}
|
|
}
|
|
|
|
// (radius + 1)^2 itself doesn't fit 32 bits here.
|
|
TEST(Imgproc_StackBlur, full_range_huge_kernel)
|
|
{
|
|
const int kernel = 92681;
|
|
for (int depth : {CV_16U, CV_16S})
|
|
{
|
|
const double value = depth == CV_16U ? 65535 : -32768;
|
|
Mat constant(1, kernel + 2, depth, Scalar::all(value)), dst;
|
|
stackBlur(constant, dst, Size(kernel, 1));
|
|
EXPECT_LE(cvtest::norm(constant, dst, NORM_INF), 1.) << "depth=" << depth;
|
|
}
|
|
}
|
|
|
|
class ScopedThreadsGuard
|
|
{
|
|
public:
|
|
ScopedThreadsGuard() : old_threads(getNumThreads()) {}
|
|
~ScopedThreadsGuard() { setNumThreads(old_threads); }
|
|
void set(int n) { setNumThreads(n); }
|
|
private:
|
|
int old_threads;
|
|
};
|
|
|
|
// The result doesn't depend on how the image is split between threads. The width is
|
|
// odd so that some stripe leaves pixels to the scalar loop at any thread count.
|
|
TEST(Imgproc_StackBlur, full_range_thread_invariance)
|
|
{
|
|
Mat noise(120, 161, CV_16UC3), single, multi;
|
|
randu(noise, Scalar::all(0), Scalar::all(65535));
|
|
|
|
ScopedThreadsGuard threadsGuard;
|
|
const int threads = std::max(2, getNumThreads());
|
|
|
|
threadsGuard.set(1);
|
|
stackBlur(noise, single, Size(5, 5));
|
|
threadsGuard.set(threads);
|
|
stackBlur(noise, multi, Size(5, 5));
|
|
|
|
EXPECT_EQ(cvtest::norm(single, multi, NORM_INF), 0.);
|
|
}
|
|
}
|
|
}
|