[Olena-patches] olena-2.0-40-g0885c7d Add: better integration of the tv diffusion term. cleaning. fix image copy.

--- milena/mln/inpainting/lpde.hh | 172 ++++++++++++++++++----------------------- 1 files changed, 76 insertions(+), 96 deletions(-) diff --git a/milena/mln/inpainting/lpde.hh b/milena/mln/inpainting/lpde.hh index f38b7a3..2eeca78 100644 --- a/milena/mln/inpainting/lpde.hh +++ b/milena/mln/inpainting/lpde.hh @@ -1,3 +1,6 @@ +#ifndef LPDE_HH_ +# define LPDE_HH_ + #include <algorithm> #include <vector> #include <limits> @@ -67,6 +70,8 @@ namespace mln this->dt = dt; this->dx = dx; this->t_f = t_f; + + this->w = std::vector<float_type>(6, 0.f); } template <typename T> @@ -114,9 +119,7 @@ namespace mln I& u_next, const std::vector<lpde::float_type>& w, const M& mask, - lpde::float_type& max_diff_norm, - I& u_x, - I& u_y) + lpde::float_type& max_diff_norm) { typedef lpde::float_type float_type; typedef mln_value(I) T; @@ -143,86 +146,78 @@ namespace mln mln_pixter(I) p_next(u_next); mln_pixter(const M) p_m(mask); - mln_pixter(I) pu_x(u_x); - mln_pixter(I) pu_y(u_y); - mln_qixter(const I, const window2d) q(p, full3x3); float_type sse = 0; - for (p.start(), p_next.start(), p_m.start(), - pu_x.start(), pu_y.start(); + for (p.start(), p_next.start(), p_m.start(); p.is_valid(); - p.next(), p_next.next(), p_m.next(), - pu_x.next(), pu_y.next()) + p.next(), p_next.next(), p_m.next()) { if (p_m.val()) - { - const T old = p.val(); - - if (fabs(w[1]) > activation_threshold) - p_next.val() = p.val() * w[1]; - - T p_derivative[derivative_kernels_size]; - - for (unsigned i = 0; i < derivative_kernels_size; ++i) - p_derivative[i] = literal::zero; - - unsigned k = 0; - for_all (q) { - for (unsigned i = 0; i < derivative_kernels_size; ++i) - { - p_derivative[i] += q.val() * derivative_kernel[i][k]; - } - ++k; - } - - const T& p_x = p_derivative[0]; - const T& p_y = p_derivative[1]; - const T& p_xx = p_derivative[2]; - const T& p_yy = p_derivative[3]; - const T& p_xy = p_derivative[4]; - - const T p_x2 = components_product(p_x, p_x); - const T p_y2 = components_product(p_y, p_y); - const T p_xx2 = components_product(p_xx, p_xx); - const T p_yy2 = components_product(p_yy, p_yy); - const T p_xy2 = components_product(p_xy, p_xy); + const T old = p.val(); - pu_x.val() = normalize(p_x); - pu_y.val() = normalize(p_y); + if (fabs(w[1]) > activation_threshold) + p_next.val() = p.val() * w[1]; - //||∇u||^2 - if (fabs(w[2]) > activation_threshold) - p_next.val() += w[2] * (p_x2 + p_y2); + T p_derivative[derivative_kernels_size]; - // laplacian - if (fabs(w[3]) > activation_threshold) - p_next.val() += w[3] * (p_xx + p_yy); - - // tr(H^2(u)) - if (fabs(w[4]) > activation_threshold) - p_next.val() += w[4] * (p_xx2 + 2 * p_xy2 + p_yy2); - - // (∇u)^T H_u (∇u) - if (fabs(w[5]) > activation_threshold) - p_next.val() += w[5] * (components_product(p_x2, p_xx) - + 2 * components_product(components_product(p_x, p_y), p_xy) - + components_product(p_y2, p_yy)); - - float_type diff = norm::l2(old - p_next.val()); - sse += diff; + for (unsigned i = 0; i < derivative_kernels_size; ++i) + p_derivative[i] = literal::zero; - if (diff > max_diff_norm) + unsigned k = 0; + for_all (q) { - max_diff_norm = diff; + for (unsigned i = 0; i < derivative_kernels_size; ++i) + { + p_derivative[i] += q.val() * derivative_kernel[i][k]; + } + ++k; } - } + + const T& p_x = p_derivative[0]; + const T& p_y = p_derivative[1]; + const T& p_xx = p_derivative[2]; + const T& p_yy = p_derivative[3]; + const T& p_xy = p_derivative[4]; + + const T p_x2 = components_product(p_x, p_x); + const T p_y2 = components_product(p_y, p_y); + const T p_xx2 = components_product(p_xx, p_xx); + const T p_yy2 = components_product(p_yy, p_yy); + const T p_xy2 = components_product(p_xy, p_xy); + + // //||∇u||^2 + if (fabs(w[2]) > activation_threshold) + p_next.val() += w[2] * (p_x2 + p_y2); + + // // laplacian + if (fabs(w[3]) > activation_threshold) + p_next.val() += w[3] * (p_xx + p_yy); + + // // tr(H^2(u)) + if (fabs(w[4]) > activation_threshold) + p_next.val() += w[4] * (p_xx2 + 2 * p_xy2 + p_yy2); + + // // (∇u)^T H_u (∇u) + if (fabs(w[5]) > activation_threshold) + p_next.val() += w[5] * (components_product(p_x2, p_xx) + + 2 * components_product(components_product(p_x, p_y), p_xy) + + components_product(p_y2, p_yy)); + + internal::interval(p_next.val(), 0.0, 1.0); + + float_type diff = norm::l2(old - p_next.val()); + sse += diff; + + if (diff > max_diff_norm) + { + max_diff_norm = diff; + } + } } - - inpainting::tv(u_next, u_x, u_y, mask, max_diff_norm); - + return sse; } @@ -230,14 +225,8 @@ namespace mln void lpde::operator()(I<T>& src, const I<bool>& mask) { - // mean_diffusion mean_diff; - // mean_diff(src, mask); - - // fast_isotropic_diffusion fast_iso; - // fast_iso(src, mask); - I<T>& u_ = src; - I<T> u_next_(src); + I<T> u_next_ = duplicate(src); extension::adjust_duplicate(u_, 1); extension::adjust_duplicate(u_next_, 1); @@ -245,36 +234,24 @@ namespace mln I<T>* u = &u_; I<T>* u_next = &u_next_; - float_type sse = std::numeric_limits<float_type>::max() / 2.0; - - float_type sse_old = std::numeric_limits<float_type>::max(); - float_type max_diff_norm = std::numeric_limits<float_type>::max(); - float_type min_max_diff_norm = std::numeric_limits<float_type>::max(); + inpainting::tv_step<I, T> tv(src, mask); - I<T> u_x(u_.domain()); - I<T> u_y(u_.domain()); - - float t = 0; + unsigned t = 0; while (t < this->t_f && max_diff_norm > CONVERGENCE_THRESHOLD - && !std::isnan(max_diff_norm) - ) + && !std::isnan(max_diff_norm)) { - max_diff_norm = 0; - sse_old = sse; - - sse = diffuse(*u, *u_next, w, mask, max_diff_norm, u_x, u_y); - - if (max_diff_norm < min_max_diff_norm) - min_max_diff_norm = max_diff_norm; - + float sse = diffuse(*u, *u_next, w, mask, max_diff_norm); + + tv(*u_next); + std::swap(u, u_next); - t += this->dt; + ++t; } std::cout << t << "\t"; @@ -282,8 +259,11 @@ namespace mln void lpde::set(const std::vector<float_type>& w) { - this->w = w; + for (unsigned i = 0; i < std::min(w.size(), this->w.size()); ++i) + this->w[i] = w[i]; } } /* mln::inpainting */ } /* mln */ + +#endif -- 1.7.2.5