// Copyright (C) 2009 Davis E. King (davis@dlib.net) // License: Boost Software License See LICENSE.txt for the full license. #ifndef DLIB_SURf_H_ #define DLIB_SURf_H_ #include "surf_abstract.h" #include "hessian_pyramid.h" #include "../matrix.h" namespace dlib { // ---------------------------------------------------------------------------------------- struct surf_point { interest_point p; matrix<double,64,1> des; double angle; }; // ---------------------------------------------------------------------------------------- inline void serialize( const surf_point& item, std::ostream& out ) { try { serialize(item.p,out); serialize(item.des,out); serialize(item.angle,out); } catch (serialization_error& e) { throw serialization_error(e.info + "\n while serializing object of type surf_point"); } } // ---------------------------------------------------------------------------------------- inline void deserialize( surf_point& item, std::istream& in ) { try { deserialize(item.p,in); deserialize(item.des,in); deserialize(item.angle,in); } catch (serialization_error& e) { throw serialization_error(e.info + "\n while deserializing object of type surf_point"); } } // ---------------------------------------------------------------------------------------- inline double gaussian (double x, double y, double sig) { DLIB_ASSERT(sig > 0, "\tdouble gaussian()" << "\n\t sig must be bigger than 0" << "\n\t sig: " << sig ); const double sqrt_2_pi = 2.5066282746310002416123552393401041626930; return 1.0/(sig*sqrt_2_pi) * std::exp( -(x*x + y*y)/(2*sig*sig)); } // ---------------------------------------------------------------------------------------- template <typename integral_image_type, typename T> double compute_dominant_angle ( const integral_image_type& img, const dlib::vector<T,2>& center, const double& scale ) { DLIB_ASSERT(get_rect(img).contains(centered_rect(center, (unsigned long)(17*scale),(unsigned long)(17*scale))) == true && scale > 0, "\tdouble compute_dominant_angle(img, center, scale)" << "\n\tAll arguments to this function must be > 0" << "\n\t get_rect(img): " << get_rect(img) << "\n\t center: " << center << "\n\t scale: " << scale ); std::vector<double> ang; std::vector<dlib::vector<double,2> > samples; const long sc = static_cast<long>(scale+0.5); // accumulate a bunch of angle and vector samples dlib::vector<double,2> vect; for (long r = -6; r <= 6; ++r) { for (long c = -6; c <= 6; ++c) { if (r*r + c*c < 36) { // compute a Gaussian weighted gradient and the gradient's angle. const double gauss = gaussian(c,r, 2.5); vect.x() = gauss*haar_x(img, sc*point(c,r)+center, 4*sc); vect.y() = gauss*haar_y(img, sc*point(c,r)+center, 4*sc); samples.push_back(vect); ang.push_back(atan2(vect.y(), vect.x())); } } } // now find the dominant direction double max_length = 0; double best_ang = 0; // look at a bunch of pie shaped slices of a circle const long slices = 45; const double ang_step = (2*pi)/slices; for (long ang_i = 0; ang_i < slices; ++ang_i) { // compute the bounding angles double ang1 = ang_step*ang_i - pi; double ang2 = ang1 + pi/3; // compute sum of all vectors that are within the above two angles vect.x() = 0; vect.y() = 0; for (unsigned long i = 0; i < ang.size(); ++i) { if (ang1 <= ang[i] && ang[i] <= ang2) { vect += samples[i]; } else if (ang2 > pi && (ang[i] >= ang1 || ang[i] <= (-2*pi+ang2))) { vect += samples[i]; } } // record the angle of the best vectors if (length_squared(vect) > max_length) { max_length = length_squared(vect); best_ang = atan2(vect.y(), vect.x()); } } return best_ang; } // ---------------------------------------------------------------------------------------- template <typename integral_image_type, typename T, typename MM, typename L> void compute_surf_descriptor ( const integral_image_type& img, const dlib::vector<T,2>& center, const double scale, const double angle, matrix<double,64,1,MM,L>& des ) { DLIB_ASSERT(get_rect(img).contains(centered_rect(center, (unsigned long)(32*scale),(unsigned long)(32*scale))) == true && scale > 0, "\tvoid compute_surf_descriptor(img, center, scale, angle)" << "\n\tAll arguments to this function must be > 0" << "\n\t get_rect(img): " << get_rect(img) << "\n\t center: " << center << "\n\t scale: " << scale ); point_rotator rot(angle); point_rotator inv_rot(-angle); const long sc = static_cast<long>(scale+0.5); long count = 0; // loop over the 4x4 grid of histogram buckets for (long r = -10; r < 10; r += 5) { for (long c = -10; c < 10; c += 5) { dlib::vector<double,2> vect, abs_vect, temp; // now loop over 25 points in this bucket and sum their features. Note // that we include 1 pixels worth of padding around the outside of each 5x5 // cell. This is to help neighboring cells interpolate their counts into // each other a little bit. for (long y = r-1; y < r+5+1; ++y) { if (y < -10 || y >= 10) continue; for (long x = c-1; x < c+5+1; ++x) { if (x < -10 || x >= 10) continue; // get the rotated point for this extraction point point p(rot(point(x,y)*scale) + center); // Give points farther from the center of the bucket a lower weight. const long center_r = r+2; const long center_c = c+2; const double weight = 1.0/(4+std::abs(center_r-y) + std::abs(center_c-x)); temp.x() = weight*haar_x(img, p, 2*sc); temp.y() = weight*haar_y(img, p, 2*sc); // rotate this vector into alignment with the surf descriptor box temp = inv_rot(temp); vect += temp; abs_vect += abs(temp); } } des(count++) = vect.x(); des(count++) = vect.y(); des(count++) = abs_vect.x(); des(count++) = abs_vect.y(); } } // Return the length normalized descriptor. Add a small number // to guard against division by zero. const double len = length(des) + 1e-7; des = des/len; } // ---------------------------------------------------------------------------------------- template <typename image_type> const std::vector<surf_point> get_surf_points ( const image_type& img, long max_points = 10000, double detection_threshold = 30.0 ) { DLIB_ASSERT(max_points > 0 && detection_threshold >= 0, "\t std::vector<surf_point> get_surf_points()" << "\n\t Invalid arguments were given to this function." << "\n\t max_points: " << max_points << "\n\t detection_threshold: " << detection_threshold ); // Figure out the proper scalar type we should use to work with these pixels. typedef typename pixel_traits<typename image_traits<image_type>::pixel_type>::basic_pixel_type bp_type; typedef typename promote<bp_type>::type working_pixel_type; // make an integral image first integral_image_generic<working_pixel_type> int_img; int_img.load(img); // now make a hessian pyramid hessian_pyramid pyr; pyr.build_pyramid(int_img, 4, 6, 2); // now get all the interest points from the hessian pyramid std::vector<interest_point> points; get_interest_points(pyr, detection_threshold, points); std::vector<surf_point> spoints; // sort all the points by how strong their detect is std::sort(points.rbegin(), points.rend()); // now extract SURF descriptors for the points surf_point sp; for (unsigned long i = 0; i < std::min((size_t)max_points,points.size()); ++i) { // ignore points that are close to the edge of the image const double border = 32; const unsigned long border_size = static_cast<unsigned long>(border*points[i].scale); if (get_rect(int_img).contains(centered_rect(points[i].center, border_size, border_size))) { sp.angle = compute_dominant_angle(int_img, points[i].center, points[i].scale); compute_surf_descriptor(int_img, points[i].center, points[i].scale, sp.angle, sp.des); sp.p = points[i]; spoints.push_back(sp); } } return spoints; } // ---------------------------------------------------------------------------------------- } #endif // DLIB_SURf_H_