2 // Copyright 2019 Olzhas Zhumabek <anonymous.from.applecity@gmail.com>
4 // Use, modification and distribution are subject to the Boost Software License,
5 // Version 1.0. (See accompanying file LICENSE_1_0.txt or copy at
6 // http://www.boost.org/LICENSE_1_0.txt)
8 #include <boost/gil/image.hpp>
9 #include <boost/gil/image_view.hpp>
10 #include <boost/gil/extension/io/png.hpp>
11 #include <boost/gil/image_processing/numeric.hpp>
12 #include <boost/gil/image_processing/harris.hpp>
13 #include <boost/gil/extension/numeric/convolve.hpp>
20 namespace gil = boost::gil;
22 // some images might produce artifacts
23 // when converted to grayscale,
24 // which was previously observed on
25 // canny edge detector for test input
26 // used for this example
27 gil::gray8_image_t to_grayscale(gil::rgb8_view_t original)
29 gil::gray8_image_t output_image(original.dimensions());
30 auto output = gil::view(output_image);
31 constexpr double max_channel_intensity = (std::numeric_limits<std::uint8_t>::max)();
32 for (long int y = 0; y < original.height(); ++y) {
33 for (long int x = 0; x < original.width(); ++x) {
34 // scale the values into range [0, 1] and calculate linear intensity
35 double red_intensity = original(x, y).at(std::integral_constant<int, 0>{})
36 / max_channel_intensity;
37 double green_intensity = original(x, y).at(std::integral_constant<int, 1>{})
38 / max_channel_intensity;
39 double blue_intensity = original(x, y).at(std::integral_constant<int, 2>{})
40 / max_channel_intensity;
41 auto linear_luminosity = 0.2126 * red_intensity
42 + 0.7152 * green_intensity
43 + 0.0722 * blue_intensity;
45 // perform gamma adjustment
46 double gamma_compressed_luminosity = 0;
47 if (linear_luminosity < 0.0031308) {
48 gamma_compressed_luminosity = linear_luminosity * 12.92;
50 gamma_compressed_luminosity = 1.055 * std::pow(linear_luminosity, 1 / 2.4) - 0.055;
53 // since now it is scaled, descale it back
54 output(x, y) = gamma_compressed_luminosity * max_channel_intensity;
61 void apply_gaussian_blur(gil::gray8_view_t input_view, gil::gray8_view_t output_view)
63 constexpr static auto filterHeight = 5ull;
64 constexpr static auto filterWidth = 5ull;
65 constexpr static double filter[filterHeight][filterWidth] =
73 constexpr double factor = 1.0 / 159;
74 constexpr double bias = 0.0;
76 const auto height = input_view.height();
77 const auto width = input_view.width();
78 for (long x = 0; x < width; ++x) {
79 for (long y = 0; y < height; ++y) {
80 double intensity = 0.0;
81 for (size_t filter_y = 0; filter_y < filterHeight; ++filter_y) {
82 for (size_t filter_x = 0; filter_x < filterWidth; ++filter_x) {
83 int image_x = x - filterWidth / 2 + filter_x;
84 int image_y = y - filterHeight / 2 + filter_y;
85 if (image_x >= input_view.width() || image_x < 0
86 || image_y >= input_view.height() || image_y < 0) {
89 auto& pixel = input_view(image_x, image_y);
90 intensity += pixel.at(std::integral_constant<int, 0>{})
91 * filter[filter_y][filter_x];
94 auto& pixel = output_view(gil::point_t(x, y));
95 pixel = (std::min)((std::max)(int(factor * intensity + bias), 0), 255);
101 std::vector<gil::point_t> suppress(
102 gil::gray32f_view_t harris_response,
103 double harris_response_threshold)
105 std::vector<gil::point_t> corner_points;
106 for (gil::gray32f_view_t::coord_t y = 1; y < harris_response.height() - 1; ++y)
108 for (gil::gray32f_view_t::coord_t x = 1; x < harris_response.width() - 1; ++x)
110 auto value = [](gil::gray32f_pixel_t pixel) {
111 return pixel.at(std::integral_constant<int, 0>{});
114 value(harris_response(x - 1, y - 1)),
115 value(harris_response(x, y - 1)),
116 value(harris_response(x + 1, y - 1)),
117 value(harris_response(x - 1, y)),
118 value(harris_response(x, y)),
119 value(harris_response(x + 1, y)),
120 value(harris_response(x - 1, y + 1)),
121 value(harris_response(x, y + 1)),
122 value(harris_response(x + 1, y + 1))
125 auto maxima = *std::max_element(
128 [](double lhs, double rhs)
134 if (maxima == value(harris_response(x, y))
135 && std::count(values, values + 9, maxima) == 1
136 && maxima >= harris_response_threshold)
138 corner_points.emplace_back(x, y);
143 return corner_points;
146 int main(int argc, char* argv[])
150 std::cout << "usage: " << argv[0] << " <input.png> <odd-window-size>"
151 " <discrimination-constant> <harris-response-threshold> <output.png>\n";
155 std::size_t window_size = std::stoul(argv[2]);
156 double discrimnation_constant = std::stof(argv[3]);
157 long harris_response_threshold = std::stol(argv[4]);
159 gil::rgb8_image_t input_image;
161 gil::read_image(argv[1], input_image, gil::png_tag{});
163 auto input_view = gil::view(input_image);
164 auto grayscaled = to_grayscale(input_view);
165 gil::gray8_image_t smoothed_image(grayscaled.dimensions());
166 auto smoothed = gil::view(smoothed_image);
167 apply_gaussian_blur(gil::view(grayscaled), smoothed);
168 gil::gray16s_image_t x_gradient_image(grayscaled.dimensions());
169 gil::gray16s_image_t y_gradient_image(grayscaled.dimensions());
171 auto x_gradient = gil::view(x_gradient_image);
172 auto y_gradient = gil::view(y_gradient_image);
173 auto scharr_x = gil::generate_dx_scharr();
174 gil::detail::convolve_2d(smoothed, scharr_x, x_gradient);
175 auto scharr_y = gil::generate_dy_scharr();
176 gil::detail::convolve_2d(smoothed, scharr_y, y_gradient);
178 gil::gray32f_image_t m11(x_gradient.dimensions());
179 gil::gray32f_image_t m12_21(x_gradient.dimensions());
180 gil::gray32f_image_t m22(x_gradient.dimensions());
181 gil::compute_tensor_entries(
189 gil::gray32f_image_t harris_response(x_gradient.dimensions());
190 auto gaussian_kernel = gil::generate_gaussian_kernel(window_size, 0.84089642);
191 gil::compute_harris_responses(
196 discrimnation_constant,
197 gil::view(harris_response)
200 auto corner_points = suppress(gil::view(harris_response), harris_response_threshold);
201 for (auto point: corner_points)
203 input_view(point) = gil::rgb8_pixel_t(0, 0, 0);
204 input_view(point).at(std::integral_constant<int, 1>{}) = 255;
206 gil::write_view(argv[5], input_view, gil::png_tag{});