// Copyright (C) 2008 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#ifndef DLIB_EDGE_DETECTOr_
#define DLIB_EDGE_DETECTOr_
#include "edge_detector_abstract.h"
#include "../pixel.h"
#include "../array2d.h"
#include "../geometry.h"
#include <vector>
#include "../image_keypoint/build_separable_poly_filters.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
template <
typename T
>
inline char edge_orientation (
const T& x_,
const T& y_
)
{
// if this is a perfectly horizontal gradient then return right away
if (x_ == 0)
{
return '|';
}
else if (y_ == 0) // if this is a perfectly vertical gradient then return right away
{
return '-';
}
// Promote x so that when we multiply by 128 later we know overflow won't happen.
typedef typename promote<T>::type type;
type x = x_;
type y = y_;
if (x < 0)
{
x = -x;
if (y < 0)
{
y = -y;
x *= 128;
const type temp = x/y;
if (temp > 309)
return '-';
else if (temp > 53)
return '/';
else
return '|';
}
else
{
x *= 128;
const type temp = x/y;
if (temp > 309)
return '-';
else if (temp > 53)
return '\\';
else
return '|';
}
}
else
{
if (y < 0)
{
y = -y;
x *= 128;
const type temp = x/y;
if (temp > 309)
return '-';
else if (temp > 53)
return '\\';
else
return '|';
}
else
{
x *= 128;
const type temp = x/y;
if (temp > 309)
return '-';
else if (temp > 53)
return '/';
else
return '|';
}
}
}
// ----------------------------------------------------------------------------------------
template <
typename in_image_type,
typename out_image_type
>
void sobel_edge_detector (
const in_image_type& in_img_,
out_image_type& horz_,
out_image_type& vert_
)
{
typedef typename image_traits<out_image_type>::pixel_type pixel_type;
COMPILE_TIME_ASSERT(pixel_traits<pixel_type>::is_unsigned == false);
DLIB_ASSERT( !is_same_object(in_img_,horz_) && !is_same_object(in_img_,vert_) &&
!is_same_object(horz_,vert_),
"\tvoid sobel_edge_detector(in_img_, horz_, vert_)"
<< "\n\t You can't give the same image as more than one argument"
<< "\n\t is_same_object(in_img_,horz_): " << is_same_object(in_img_,horz_)
<< "\n\t is_same_object(in_img_,vert_): " << is_same_object(in_img_,vert_)
<< "\n\t is_same_object(horz_,vert_): " << is_same_object(horz_,vert_)
);
const int vert_filter[3][3] = {{-1,-2,-1},
{0,0,0},
{1,2,1}};
const int horz_filter[3][3] = { {-1,0,1},
{-2,0,2},
{-1,0,1}};
const long M = 3;
const long N = 3;
const_image_view<in_image_type> in_img(in_img_);
image_view<out_image_type> horz(horz_);
image_view<out_image_type> vert(vert_);
horz.set_size(in_img.nr(),in_img.nc());
vert.set_size(in_img.nr(),in_img.nc());
assign_border_pixels(horz,1,1,0);
assign_border_pixels(vert,1,1,0);
// figure out the range that we should apply the filter to
const long first_row = M/2;
const long first_col = N/2;
const long last_row = in_img.nr() - M/2;
const long last_col = in_img.nc() - N/2;
// apply the filter to the image
for (long r = first_row; r < last_row; ++r)
{
for (long c = first_col; c < last_col; ++c)
{
typedef typename pixel_traits<typename image_traits<in_image_type>::pixel_type>::basic_pixel_type bp_type;
typename promote<bp_type>::type p, horz_temp, vert_temp;
horz_temp = 0;
vert_temp = 0;
for (long m = 0; m < M; ++m)
{
for (long n = 0; n < N; ++n)
{
// pull out the current pixel and put it into p
p = get_pixel_intensity(in_img[r-M/2+m][c-N/2+n]);
horz_temp += p*horz_filter[m][n];
vert_temp += p*vert_filter[m][n];
}
}
assign_pixel(horz[r][c] , horz_temp);
assign_pixel(vert[r][c] , vert_temp);
}
}
}
// ----------------------------------------------------------------------------------------
namespace impl
{
template <typename T>
typename promote<T>::type square (const T& a)
{
return static_cast<T>(a)*static_cast<T>(a);
}
}
template <
typename in_image_type,
typename out_image_type
>
void suppress_non_maximum_edges (
const in_image_type& horz_,
const in_image_type& vert_,
out_image_type& out_img_
)
{
const_image_view<in_image_type> horz(horz_);
const_image_view<in_image_type> vert(vert_);
image_view<out_image_type> out_img(out_img_);
COMPILE_TIME_ASSERT(is_signed_type<typename image_traits<in_image_type>::pixel_type>::value);
DLIB_ASSERT( horz.nr() == vert.nr() && horz.nc() == vert.nc(),
"\tvoid suppress_non_maximum_edges(horz, vert, out_img)"
<< "\n\tYou have to give horz and vert gradient images that are the same size"
<< "\n\thorz.nr(): " << horz.nr()
<< "\n\thorz.nc(): " << horz.nc()
<< "\n\tvert.nr(): " << vert.nr()
<< "\n\tvert.nc(): " << vert.nc()
);
DLIB_ASSERT( !is_same_object(out_img_,horz_) && !is_same_object(out_img_,vert_),
"\tvoid suppress_non_maximum_edges(horz_, vert_, out_img_)"
<< "\n\t out_img can't be the same as one of the input images."
<< "\n\t is_same_object(out_img_,horz_): " << is_same_object(out_img_,horz_)
<< "\n\t is_same_object(out_img_,vert_): " << is_same_object(out_img_,vert_)
);
using std::min;
using std::abs;
// if there isn't any input image then don't do anything
if (horz.size() == 0)
{
out_img.clear();
return;
}
out_img.set_size(horz.nr(),horz.nc());
zero_border_pixels(out_img,1,1);
// now do non maximum suppression while we copy the
const long M = 3;
const long N = 3;
// figure out the range that we should apply the filter to
const long first_row = M/2;
const long first_col = N/2;
const long last_row = horz.nr() - M/2;
const long last_col = horz.nc() - N/2;
// apply the filter to the image
for (long r = first_row; r < last_row; ++r)
{
for (long c = first_col; c < last_col; ++c)
{
typedef typename promote<typename image_traits<in_image_type>::pixel_type>::type T;
const T y = horz[r][c];
const T x = vert[r][c];
using impl::square;
const T val = square(horz[r][c]) + square(vert[r][c]);
const char ori = edge_orientation(x,y);
const unsigned char zero = 0;
switch (ori)
{
case '-':
if (square(horz[r-1][c])+square(vert[r-1][c]) > val || square(horz[r+1][c]) + square(vert[r+1][c]) > val)
assign_pixel(out_img[r][c] , zero);
else
assign_pixel(out_img[r][c] , std::sqrt((double)val));
break;
case '|':
if (square(horz[r][c-1]) + square(vert[r][c-1]) > val || square(horz[r][c+1]) + square(vert[r][c+1]) > val)
assign_pixel(out_img[r][c] , zero);
else
assign_pixel(out_img[r][c] , std::sqrt((double)val));
break;
case '/':
if (square(horz[r-1][c-1]) + square(vert[r-1][c-1]) > val || square(horz[r+1][c+1]) + square(vert[r+1][c+1]) > val)
assign_pixel(out_img[r][c] , zero);
else
assign_pixel(out_img[r][c] , std::sqrt((double)val));
break;
case '\\':
if (square(horz[r+1][c-1]) + square(vert[r+1][c-1]) > val || square(horz[r-1][c+1]) + square(vert[r-1][c+1]) > val)
assign_pixel(out_img[r][c] , zero);
else
assign_pixel(out_img[r][c] , std::sqrt((double)val));
break;
}
}
}
}
// ----------------------------------------------------------------------------------------
template <
typename image_type
>
void normalize_image_gradients (
image_type& img1_,
image_type& img2_
)
{
image_view<image_type> img1(img1_);
image_view<image_type> img2(img2_);
using pixel_type = typename image_traits<image_type>::pixel_type;
static_assert(std::is_same<pixel_type,float>::value ||
std::is_same<pixel_type,double>::value ||
std::is_same<pixel_type,long double>::value,
"normalize_image_gradients() requires the input images to use floating point pixel types.");
DLIB_CASSERT(img1.nr() == img2.nr());
DLIB_CASSERT(img1.nc() == img2.nc());
// normalize all the gradients
for (long r = 0; r < img1.nr(); ++r)
{
for (long c = 0; c < img1.nc(); ++c)
{
if (img1[r][c] != 0 || img2[r][c] != 0)
{
double len = std::sqrt(img1[r][c]*img1[r][c] + img2[r][c]*img2[r][c]);
img1[r][c] /= len;
img2[r][c] /= len;
}
}
}
}
// ----------------------------------------------------------------------------------------
template <
typename image_type
>
std::vector<point> remove_incoherent_edge_pixels (
const std::vector<point>& line,
const image_type& horz_gradient_,
const image_type& vert_gradient_,
double angle_threshold
)
{
const_image_view<image_type> horz_gradient(horz_gradient_);
const_image_view<image_type> vert_gradient(vert_gradient_);
DLIB_CASSERT(horz_gradient.nr() == vert_gradient.nr());
DLIB_CASSERT(horz_gradient.nc() == vert_gradient.nc());
DLIB_CASSERT(angle_threshold >= 0);
#ifdef ENABLE_ASSERTS
for (auto& p : line)
DLIB_ASSERT(get_rect(horz_gradient).contains(p), "All line points must be inside the given images.");
#endif
// We make sure that each vector is within this threshold of the mean vector. So
// to make sure they are pairwise within the user supplied angel threshold we need
// to divide by 2 before we proceed.
angle_threshold /= 2;
const double dotthresh = std::cos(angle_threshold*pi/180);
// find the average gradient on this line
dpoint avg;
for (auto p : line)
avg += dpoint(horz_gradient[p.y()][p.x()], vert_gradient[p.y()][p.x()]);
dpoint ref = avg.normalize();
// now iterate a few times and find the most common average gradient.
for (int i = 0; i < 10; ++i)
{
avg = dpoint();
for (auto p : line)
{
const dpoint v(horz_gradient[p.y()][p.x()], vert_gradient[p.y()][p.x()]);
const double dp = ref.dot(v);
if (dp > dotthresh)
avg += v;
else if (-dp > dotthresh)
avg -= v;
}
ref = avg.normalize();
}
// now remove all the points that deviate from the average gradient too much.
std::vector<point> newpixels;
for (auto p : line)
{
dpoint v(horz_gradient[p.y()][p.x()], vert_gradient[p.y()][p.x()]);
if (std::abs(ref.dot(v)) > dotthresh)
newpixels.push_back(p);
}
return newpixels;
}
template <
typename image_type
>
std::vector<std::vector<point>> remove_incoherent_edge_pixels (
const std::vector<std::vector<point>>& line_pixels,
const image_type& horz_gradient_,
const image_type& vert_gradient_,
const double angle_threshold
)
{
std::vector<std::vector<point>> temp;
temp.reserve(line_pixels.size());
for (auto& line : line_pixels)
temp.emplace_back(remove_incoherent_edge_pixels(line, horz_gradient_, vert_gradient_, angle_threshold));
return temp;
}
// ----------------------------------------------------------------------------------------
class image_gradients
{
public:
image_gradients (
) : image_gradients(1) {}
image_gradients (
long scale
) : the_scale(scale)
{
DLIB_CASSERT(scale >= 1);
scale = 2*scale+1;
auto dfilters = build_separable_poly_filters(2,scale);
DLIB_CASSERT(dfilters[1].size() == 1);
DLIB_CASSERT(dfilters[2].size() == 1);
DLIB_CASSERT(dfilters[3].size() == 1);
DLIB_CASSERT(dfilters[4].size() == 1);
DLIB_CASSERT(dfilters[5].size() == 1);
filter_x.first = matrix_cast<float>(dfilters[1][0].first);
filter_x.second = matrix_cast<float>(dfilters[1][0].second);
filter_y.first = matrix_cast<float>(dfilters[2][0].first);
filter_y.second = matrix_cast<float>(dfilters[2][0].second);
// We multiply by 2 so that the filter gives the gradient rather than the x^2
// polynomial coefficient.
filter_xx.first = 2*matrix_cast<float>(dfilters[3][0].first);
filter_xx.second = matrix_cast<float>(dfilters[3][0].second);
filter_xy.first = matrix_cast<float>(dfilters[4][0].first);
filter_xy.second = matrix_cast<float>(dfilters[4][0].second);
// We multiply by 2 so that the filter gives the gradient rather than the y^2
// polynomial coefficient.
filter_yy.first = 2*matrix_cast<float>(dfilters[5][0].first);
filter_yy.second = matrix_cast<float>(dfilters[5][0].second);
}
long get_scale() const { return the_scale; }
template <
typename in_image_type,
typename out_image_type
>
rectangle gradient_x(
const in_image_type& img,
out_image_type& out
) const
{
return spatially_filter_image_separable(img, out, filter_x.second, filter_x.first);
}
template <
typename in_image_type,
typename out_image_type
>
rectangle gradient_y(
const in_image_type& img,
out_image_type& out
) const
{
return spatially_filter_image_separable(img, out, filter_y.second, filter_y.first);
}
template <
typename in_image_type,
typename out_image_type
>
rectangle gradient_xx(
const in_image_type& img,
out_image_type& out
) const
{
return spatially_filter_image_separable(img, out, filter_xx.second, filter_xx.first);
}
template <
typename in_image_type,
typename out_image_type
>
rectangle gradient_xy(
const in_image_type& img,
out_image_type& out
) const
{
return spatially_filter_image_separable(img, out, filter_xy.second, filter_xy.first);
}
template <
typename in_image_type,
typename out_image_type
>
rectangle gradient_yy(
const in_image_type& img,
out_image_type& out
) const
{
return spatially_filter_image_separable(img, out, filter_yy.second, filter_yy.first);
}
matrix<float> get_x_filter() const { return filter_x.first*trans(filter_x.second); }
matrix<float> get_y_filter() const { return filter_y.first*trans(filter_y.second); }
matrix<float> get_xx_filter() const { return filter_xx.first*trans(filter_xx.second); }
matrix<float> get_xy_filter() const { return filter_xy.first*trans(filter_xy.second); }
matrix<float> get_yy_filter() const { return filter_yy.first*trans(filter_yy.second); }
private:
std::pair<matrix<float,0,1>,matrix<float,0,1>> filter_x;
std::pair<matrix<float,0,1>,matrix<float,0,1>> filter_y;
std::pair<matrix<float,0,1>,matrix<float,0,1>> filter_xx;
std::pair<matrix<float,0,1>,matrix<float,0,1>> filter_xy;
std::pair<matrix<float,0,1>,matrix<float,0,1>> filter_yy;
long the_scale;
};
// ----------------------------------------------------------------------------------------
namespace impl
{
template <
typename in_image_type,
typename out_image_type
>
void find_lines(
const in_image_type& xx_,
const in_image_type& xy_,
const in_image_type& yy_,
out_image_type& horz_,
out_image_type& vert_,
double positive_if_should_find_dark_lines
)
{
typedef typename image_traits<out_image_type>::pixel_type out_pixel_type;
static_assert(std::is_same<float,out_pixel_type>::value || std::is_same<double,out_pixel_type>::value,
"Output images must contain either float or double valued pixels");
const_image_view<in_image_type> xx(xx_);
const_image_view<in_image_type> xy(xy_);
const_image_view<in_image_type> yy(yy_);
DLIB_CASSERT(xx.nr() == xy.nr());
DLIB_CASSERT(xx.nr() == yy.nr());
DLIB_CASSERT(xx.nc() == xy.nc());
DLIB_CASSERT(xx.nc() == yy.nc());
image_view<out_image_type> x(horz_);
image_view<out_image_type> y(vert_);
x.set_size(xx.nr(), xx.nc());
y.set_size(xx.nr(), xx.nc());
// store the max eigenvalue into xy and then the associated eigen vector into [xx,yy]
for (long r = 0; r < xx.nr(); ++r)
{
for (long c = 0; c < xx.nc(); ++c)
{
// negate to that lambda will be the *minimum* eigenvalue
double w1 = positive_if_should_find_dark_lines*xx[r][c]/2.0;
double w2 = positive_if_should_find_dark_lines*yy[r][c]/2.0;
double w3 = positive_if_should_find_dark_lines*xy[r][c];
auto lambda = w1 + w2 + std::sqrt((w1-w2)*(w1-w2) + w3*w3);
if (lambda < 0)
lambda = 0;
if (2*w1!=lambda)
{
x[r][c] = -w3/(2*w1-lambda);
y[r][c] = 1;
double norm = std::sqrt(x[r][c]*x[r][c] + y[r][c]*y[r][c]);
x[r][c] *= lambda/norm;
y[r][c] *= lambda/norm;
}
else
{
x[r][c] = lambda;
y[r][c] = 0;
}
}
}
}
}
template <
typename in_image_type,
typename out_image_type
>
void find_bright_lines(
const in_image_type& xx,
const in_image_type& xy,
const in_image_type& yy,
out_image_type& horz,
out_image_type& vert
)
{
impl::find_lines(xx,xy,yy,horz,vert,-1);
}
template <
typename in_image_type,
typename out_image_type
>
void find_dark_lines(
const in_image_type& xx,
const in_image_type& xy,
const in_image_type& yy,
out_image_type& horz,
out_image_type& vert
)
{
impl::find_lines(xx,xy,yy,horz,vert,+1);
}
// ----------------------------------------------------------------------------------------
template <
typename in_image_type,
typename out_image_type
>
void find_bright_keypoints(
const in_image_type& xx_,
const in_image_type& xy_,
const in_image_type& yy_,
out_image_type& saliency_
)
{
typedef typename image_traits<out_image_type>::pixel_type out_pixel_type;
static_assert(std::is_same<float,out_pixel_type>::value || std::is_same<double,out_pixel_type>::value,
"Output images must contain either float or double valued pixels");
const_image_view<in_image_type> xx(xx_);
const_image_view<in_image_type> xy(xy_);
const_image_view<in_image_type> yy(yy_);
DLIB_CASSERT(xx.nr() == xy.nr());
DLIB_CASSERT(xx.nr() == yy.nr());
DLIB_CASSERT(xx.nc() == xy.nc());
DLIB_CASSERT(xx.nc() == yy.nc());
image_view<out_image_type> saliency(saliency_);
saliency.set_size(xx.nr(), xx.nc());
for (long r = 0; r < xx.nr(); ++r)
{
for (long c = 0; c < xx.nc(); ++c)
{
matrix<double,2,2> tmp;
tmp = xx[r][c], xy[r][c],
xy[r][c], yy[r][c];
matrix<double,2,1> e = real_eigenvalues(tmp);
saliency[r][c] = prod(upperbound(e,0));
}
}
}
// ----------------------------------------------------------------------------------------
template <
typename in_image_type,
typename out_image_type
>
void find_dark_keypoints(
const in_image_type& xx_,
const in_image_type& xy_,
const in_image_type& yy_,
out_image_type& saliency_
)
{
typedef typename image_traits<out_image_type>::pixel_type out_pixel_type;
static_assert(std::is_same<float,out_pixel_type>::value || std::is_same<double,out_pixel_type>::value,
"Output images must contain either float or double valued pixels");
const_image_view<in_image_type> xx(xx_);
const_image_view<in_image_type> xy(xy_);
const_image_view<in_image_type> yy(yy_);
DLIB_CASSERT(xx.nr() == xy.nr());
DLIB_CASSERT(xx.nr() == yy.nr());
DLIB_CASSERT(xx.nc() == xy.nc());
DLIB_CASSERT(xx.nc() == yy.nc());
image_view<out_image_type> saliency(saliency_);
saliency.set_size(xx.nr(), xx.nc());
for (long r = 0; r < xx.nr(); ++r)
{
for (long c = 0; c < xx.nc(); ++c)
{
matrix<double,2,2> tmp;
tmp = xx[r][c], xy[r][c],
xy[r][c], yy[r][c];
matrix<double,2,1> e = real_eigenvalues(tmp);
saliency[r][c] = prod(lowerbound(e,0));
}
}
}
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_EDGE_DETECTOr_