https://svn.lrde.epita.fr/svn/oln/branches/cleanup-2008/milena/sandbox Index: ChangeLog from Alexandre Abraham &lt;abraham(a)lrde.epita.fr&gt; Make the neighborhood energy count. * markov/markov.hh: . markov.hh | 45 +++++++++++++++++++++++++++++++-------------- 1 file changed, 31 insertions(+), 14 deletions(-) Index: markov/markov.hh --- markov/markov.hh (revision 2894) +++ markov/markov.hh (working copy) @@ -9,8 +9,8 @@ namespace mln { - template <typename I, typename N> - double compute_energy(const I& ima, const N& nbh, bool xi, const mln_site(I) &p) + template <typename I, typename O, typename N> + double compute_energy(const I& ima, const O& out, const N& nbh, bool xi, const mln_site(I) &p) { // Compute u(x,y) double u; @@ -25,19 +25,19 @@ mln_niter(N) n(nbh, p); for_all(n) - diff_sum += abs(ima(p) - ima(n)); + { + diff_sum += abs(out(p) - out(n)); + } - return u + 10 * diff_sum; + return u + 100 * diff_sum; } template <typename I, typename N> // I == int_u8 - mln_ch_value(I, bool) markov(const Image<I>& ima_, const Neighborhood<N> nbh_, unsigned start_temp) + mln_ch_value(I, bool) markov(const Image<I>& ima_, const Neighborhood<N>& nbh_, unsigned start_temp) { const I &ima = exact(ima_); const N &nbh = exact(nbh_); - double epsilon = 0.001; - // mln_ch_value(I, bool) out(ima.domain()); // FIXME: generalize, we might not do a binarisation mln_ch_value(I, bool) out = binarization::threshold(ima, 255 / 2); // FIXME : max @@ -47,25 +47,42 @@ Random<bool> v_random(0, 1); // mettre max et min ? Random<double> p_random(0., 1.); // idem - while (temp > epsilon) + unsigned modifications = 42; + unsigned turn = 1; + bool gradient = false; + + while (!gradient || modifications) { mln_piter(I) p(ima.domain()); + modifications = 0; for_all(p) { bool v = v_random.get(); - double u = compute_energy(ima, nbh, out(p), p); - double up = compute_energy(ima, nbh, v, p); - - double d_u = abs(up - u); + double u = compute_energy(ima, out, nbh, out(p), p); + double up = compute_energy(ima, out, nbh, v, p); + double d_u = up - u; double proba = exp(-d_u / temp); - if (d_u < 0 || (p_random.get() > proba)) + // std::cout << "Difference : " << d_u << std::endl; + + if (d_u < 0 || !gradient && (p_random.get() < proba)) + { out(p) = v; + modifications ++; + } } temp = gtemp; + std::cout << "Turn : " << turn << " Modifs : " << modifications << " Temp : " << temp << std::endl; + turn ++; + if (!gradient && !modifications) + { + std::cout << "Gradient !" << std::endl; + modifications = 1; + gradient = true; + } } return out;