int main(){ unsigned char *ro_ch,*ve_ch,*az_ch; float *red_input,*red_output,*blue_input,*blue_output,*green_input,*green_output,lambda; char fentrada[200]; int i,niter,tam,h,w; float dt; printf("---------------MENU-----------------.\n"); printf("Introduzca la direccion del archivo\n"); scanf("%s",fentrada); printf("introduzca dt \n"); scanf("%f",&dt); printf("Introduzca numero de iteraciones\n"); scanf("%d",&niter); printf("Introduzca valor para el parámetro lambda\n"); scanf("%f",&lambda); ami_read_bmp(fentrada,&ro_ch,&ve_ch,&az_ch, &w,&h); tam=w*h; red_input=(float*)malloc(tam*sizeof(float)); blue_input=(float*)malloc(tam*sizeof(float)); green_input=(float*)malloc(tam*sizeof(float)); red_output=(float*)malloc(tam*sizeof(float)); blue_output=(float*)malloc(tam*sizeof(float)); green_output=(float*)malloc(tam*sizeof(float)); for(i=0;i<tam;i++){ red_input[i]=ro_ch[i]; blue_input[i]=az_ch[i]; green_input[i]=ve_ch[i]; } perona_malik(red_input,green_input,blue_input,w,h,dt,niter,lambda); free(red_input); free(blue_input); free(green_input); free(red_output); free(blue_output); free(green_output); free(ro_ch); free(az_ch); free(ve_ch); printf("FINISH"); return 0; }
int main(int argc, char ** argv) { /// @section csv_segmentation Chan-Sandberg-Vese segmentation /// /// @subsection csv_theory Theory /// Since the routine contains too many free parameters which makes it unreasonable to place it into a separate /// function, all the code is kept in main(). Here's a rough explanation of what's Chan-Sandberg-Vese all about, /// which is based on paper @cite Getreuer2012. /// /// The Chan-Vese method seeks a contour @f$\mathcal{C}@f$ which minimizes the functional /// @f[ /// \mathcal{F}[I;\,\mathcal{C},\,c_{1},\,c_{2}]= /// \mu\mathrm{Length}(\mathcal{C})+ /// \nu\mathrm{Area}(\mathcal{C})+ /// \lambda_{1}\int_{\mathcal{C}}|I-c_{1}|^{2}\,\mathrm{d}x\mathrm{d}y+ /// \lambda_{2}\int_{\Omega\setminus\mathcal{C}}|I-c_{2}|^{2}\,\mathrm{d}x\mathrm{d}y\,, /// @f] /// where /// - the single-channel image @f$I=I(x,\,y)@f$ is defined on the region @f$\Omega=[0,\,a]\times[0,\,b]@f$; /// - regions in the integral limits, @f$\mathcal{C}@f$ and @f$\Omega\setminus\mathcal{C}@f$, /// denote the region enclosed by the contour and the region outside the contour, respectively; /// - @f$\mu(=0.5)@f$, @f$\nu(=0)@f$, @f$\lambda_{1}(=1)@f$ and @f$\lambda_{2}(=1)@f$ are free parameters, /// whereby only @f$\nu@f$ can be negative (default values in parentheses); /// - @f$c_{1}@f$ and @f$c_{2}@f$ are constants that depend on the information of the regions enclosed by and /// outside of the contour. /// /// Instead of dealing with @f$\mathcal{C}@f$ explicitly, it's custom to define a level set function @f$u(x,\,y;\,t)@f$ /// so that its zero-level iso-surface (also: zero level set) coincides with the contour: /// @f$\mathcal{C}=\{\Omega\ni(x,\,y)\,:\,u(x,\,y;\,t)=0\forall t\}@f$. This in turn leads us to a new definition /// of the functional: /// @f[ /// \mathcal{F}[I;\,u,\,c_{1},\,c_{2}] = /// \mu\left(\int_{\Omega}|\nabla H(u)|\,\mathrm{d}x\mathrm{d}y\right)^{p}+ /// \nu\int_{\Omega}H(u)\,\mathrm{d}x\mathrm{d}y+ /// \lambda_{1}\int_{\Omega}|I-c_{1}|^{2}H(u)\,\mathrm{d}x\mathrm{d}y+ /// \lambda_{2}\int_{\Omega}|I-c_{2}|^{2}(1-H(u))\,\mathrm{d}x\mathrm{d}y\,. /// @f] /// In our implementation we've picked @f$p=1@f$, so that the first integral reduces to /// @f[ /// \left.\mu\left(\int_{\Omega}|\nabla H(u)|\,\mathrm{d}x\mathrm{d}y\right)^{p}\right|_{p=1}= /// \mu\int_{\Omega}\delta(u)|\nabla u|\,\mathrm{d}x\mathrm{d}y\,, /// @f] /// where @f$H(x)@f$ and @f$\delta(x)=H'(x)@f$ are Heaviside's step and Dirac's delta functions. /// In this prescription @f$c_{1}@f$ and @f$c_{2}@f$ are now region averages and take the following form: /// @f[ /// c_{1}=\frac{\int_{\Omega}IH(u)\mathrm{d}x\mathrm{d}y}{\int_{\Omega}H(u)\,\mathrm{d}x\mathrm{d}y}\,,\quad /// c_{2}=\frac{\int_{\Omega}I(1-H(u))\mathrm{d}x\mathrm{d}y}{\int_{\Omega}(1 - H(u))\,\mathrm{d}x\mathrm{d}y}\,. /// @f] /// For practical reasons the functions are replaced by smooth/regularized versions (see regularized_heaviside() and /// regularized_delta()): /// @f[ /// H_{\epsilon}(x)=\frac{1}{2}\left[1+\frac{2}{\pi}\arctan\left(\frac{x}{\epsilon}\right)\right]\,,\quad /// \delta_{\epsilon}(x)=\frac{\epsilon}{\pi\left(\epsilon^{2}+x^{2}\right)}\,, /// @f] /// with @f$\epsilon=1@f$ by default. /// The interpretation of the functional @f$\mathcal{F}@f$ is the following: /// - the first term penalizes the length of @f$\mathcal{C}@f$; /// - the second term penalizes the area enclosed by the curve; /// - the 3rd and 4th term penalize region averages inside and outside of the contour; in other words /// it keeps track of the discrepancy between the two regions. /// /// A stationary solution to @f$\mathcal{F}@f$, or equivalently the equation of motion (e.o.m) for the contour, /// can be found by solving it with Euler-Lagrange equation, which results in /// @f[ /// u_{t} = \delta_{\epsilon}(u)\left[\mu\kappa-\nu-\lambda_{1}(I-c_{1})^{2}+\lambda_{2}(I-c_{2})^{2}\right]\,, /// @f] /// where @f$\kappa=\nabla\cdot\left(\frac{\nabla u}{|\nabla u|}\right)@f$ is curvature of @f$u@f$. /// /// If the (still 2D) image has @f$I@f$ has @f$N@f$ channels @f$\{I_{i}(x,\,y)\}_{i=1}^{N}@f$, there should still /// be a single level set @f$u@f$, which leads us the following functional: /// @f[ /// \mathcal{F}[I;\,u,\,\mathbf{c_{1}},\,\mathbf{c}_{2}]= /// \mu\int_{\Omega}|\nabla H(u)|\mathrm{d}x\mathrm{d}y+ /// \nu\int_{\Omega}H(u)\mathrm{d}x\mathrm{d}y+ /// \int_{\Omega}\frac{1}{N}\sum_{i=1}^{N}\lambda_{1}^{(i)}|I_{i}-c_{1}^{(i)}|^{2}H(u)\mathrm{d}x\mathrm{d}y+ /// \int_{\Omega}\frac{1}{N}\sum_{i=1}^{N}\lambda_{2}^{(i)}|I_{i}-c_{2}^{(i)}|^{2}(1-H(u))\mathrm{d}x\mathrm{d}y\,. /// @f] /// Variables @f$\{c_{1}^{(i)},\,c_{2}^{(i)}\}_{i=1}^{N}@f$ retain their original meaning, /// @f[ /// c_{1}^{(i)}=\frac{\int_{\Omega}I_{i}H(u)\mathrm{d}x\mathrm{d}y}{\int_{\Omega}H(u)\mathrm{d}x\mathrm{d}y}\,,\quad /// c_{2}^{(i)}=\frac{\int_{\Omega}I_{i}(1-H(u))\mathrm{d}x\mathrm{d}y}{\int_{\Omega}(1-H(u))\mathrm{d}x\mathrm{d}y} /// \quad\forall i=\{1,\,\ldots,\,N\}\,; /// @f] /// the constants @f$\{\lambda_{1}^{(i)},\,\lambda_{2}^{(i)}\}_{i=1}^{N}@f$ are defined for each channel separately. /// This implementation consider only grayscale (@f$N=1@f$) and RGB (@f$N=3@f$) images, for which @f$\lambda_{i}=1@f$ /// by default for any @f$i@f$-th channel. /// The corresponding e.o.m reads /// @f[ /// u_{t}=\delta_{\epsilon}(u)\left[\mu\kappa-\nu- /// \frac{1}{N}\sum_{i=1}^{N}\lambda_{1}^{(i)}\left(I_{i}-c_{1}^{(i)}\right)^{2}+ /// \frac{1}{N}\sum_{i=1}^{N}\lambda_{2}^{(i)}\left(I_{i}-c_{2}^{(i)}\right)^{2}\right]\,. /// @f] /// /// @subsection csv_numsch Numerical scheme /// /// Finite difference expression for the curvature @f$\kappa@f$ is explained in curvature(). The advantage of this scheme /// is that we only need nearest neighbouring points at current point while keeping the derivative centered at current point, /// whereas naive implementation would use more distant points. Since we're dealing with a finite domain and therefore /// boundaries, we don't have to "extend" the region by two pixels each direction. Instead, we just duplicate border pixels. /// /// Rest of the calculation is actually quite straightforward -- the zero level set is iteratively updated with /// @f[ /// u_{i,j}^{n+1}=u_{i,j}^{n}+\mathrm{d}t\;\delta_{\epsilon}(u_{i,j}^{n})\left[\kappa_{i,j}^{n}-\nu- /// \frac{1}{N}\sum_{k=1}^N\lambda_{1}^{(k)}\left(I_{i,j}-c_{1}^{n,(k)}\right)+ /// \frac{1}{N}\sum_{k=1}^N\lambda_{2}^{(k)}\left(I_{i,j}-c_{2}^{n,(k)}\right)\right]\,. /// @f] /// The method is inherently implicit and is implemented with ordinary matrix operations. The first term in the brackets /// has already been discussed; the second term is trivial; the final two terms are explained in region_variance() and /// variance_penalty(). /// /// There are various ways to initialize the level set, and since we're solving a differential equation, different initial /// conditions lead to different outcome. The simplest way is to let the user draw either rectangular or circular contour. /// The level set will be evaluated with @f$+1@f$'s inside the contour and with @f$-1@f$'s outside of it. /// A more optimal (here the default) contour would be checkerboard /// @f[ /// u(i,\,j;\,0)=\sin\left(\frac{\pi}{5}i\right)\sin\left(\frac{\pi}{5}j\right)\,, /// @f] /// because it converges faster to a solution (see levelset_checkerboard()). The solution is reached when the maximum number /// of iterations, @f$T_\max@f$, is reached or when @f$||u_{i,j}^{n+1}-u_{i,j}^{n}||_{2}\leqslant\delta ||\bar{I}||_{2}@f$, /// where the subscript denotes @f$L_{2}@f$-norm, @f$\delta=(10^{-3})@f$ is tolerance parameter and @f$\bar{I}@f$ is /// the intensity average in the image (averaged across the channels). /// /// @subsection csv_summary Summary /// /// The main logic described above starts with a timestep loop (look for the comment below); everything else preciding /// that is actually sugar coating just to make the program usable for anyone. /// /// If it isn't clear from above text or the code below, here is the list of variables which the user can pass as an argument /// (the default values in the parentheses): @f$\mu(=0.5)@f$, @f$\nu(=0)@f$, @f$\mathrm{d}t(=1)@f$, /// @f$\lambda_{1}^{(i)}(=1)@f$ and @f$\lambda_{1}^{(i)}(=1)@f$ @f$\forall i=1\ldots N@f$, @f$\epsilon(=1)@f$, /// @f$\delta(=10^{-3})@f$, @f$T_\max@f$(=INT_MAX), @f$N(=1\;\mbox{or}\;3)@f$ (number of channels). /// /// Other general options include: /// - object selection (-s) -- the region enclosed by the contour will be cut out and placed onto white canvas and saved; /// - region inversion (-I) -- sometimes the ROI is inverted; there's an option to circumvent that (goes with -s option); /// - video output (-V) -- see contour evolution in a video (*.avi, the same filename as the image; see VideoWriterManager); /// - overlay text (-O) -- puts overlay text (timesteps) on the video (goes with the previous option); /// - frame rate (-f) -- frame rate of the video; /// - line color (-l) -- color of the contour line (see Colors); /// - rectangular (-R) or circular (-C) contour -- lets the user draw it on the image (see InteractiveData and its subclasses); /// - grayscale image (-g) -- sometimes we just want do perform it on a black-white image, but the original source is RGB. /// /// For Perona-Malik-specific parameters @f$K@f$, @f$L@f$, @f$T@f$, see perona_malik(). /// /// @sa curvature, region_variance, variance_penalty, levelset_checkerboard, VideoWriterManager, InteractiveData, Colors, perona_malik double mu, nu, eps, tol, dt, fps, K, L, T; int max_steps; std::vector<double> lambda1, lambda2; std::string input_filename, text_position, line_color_str; bool grayscale = false, write_video = false, overlay_text = false, object_selection = false, invert = false, segment = false, rectangle_contour = false, circle_contour = false; ChanVese::TextPosition pos = ChanVese::TextPosition::TopLeft; cv::Scalar contour_color = ChanVese::Colors::blue; //-- Parse command line arguments // Negative values in multitoken are not an issue, b/c it doesn't make much sense // to use negative values for lambda1 and lambda2 try { namespace po = boost::program_options; po::options_description desc("Allowed options", get_terminal_width()); desc.add_options() ("help,h", "this message") ("input,i", po::value<std::string>(&input_filename), "input image") ("mu", po::value<double>(&mu) -> default_value(0.5), "length penalty parameter (must be positive or zero)") ("nu", po::value<double>(&nu) -> default_value(0), "area penalty parameter") ("dt", po::value<double>(&dt) -> default_value(1), "timestep") ("lambda1", po::value<std::vector<double>>(&lambda1) -> multitoken(), "penalty of variance inside the contour (default: 1's)") ("lambda2", po::value<std::vector<double>>(&lambda2) -> multitoken(), "penalty of variance outside the contour (default: 1's)") ("epsilon,e", po::value<double>(&eps) -> default_value(1), "smoothing parameter in Heaviside/delta") ("tolerance,t", po::value<double>(&tol) -> default_value(0.001), "tolerance in stopping condition") ("max-steps,N", po::value<int>(&max_steps) -> default_value(-1), "maximum nof iterations (negative means unlimited)") ("fps,f", po::value<double>(&fps) -> default_value(10), "video fps") ("overlay-pos,P", po::value<std::string>(&text_position) -> default_value("TL"), "overlay tex position; allowed only: TL, BL, TR, BR") ("line-color,l", po::value<std::string>(&line_color_str) -> default_value("blue"), "contour color (allowed only: black, white, R, G, B, Y, M, C") ("edge-coef,K", po::value<double>(&K) -> default_value(10), "coefficient for enhancing edge detection in Perona-Malik") ("laplacian-coef,L", po::value<double>(&L) -> default_value(0.25), "coefficient in the gradient FD scheme of Perona-Malik (must be [0, 1/4])") ("segment-time,T", po::value<double>(&T) -> default_value(20), "number of smoothing steps in Perona-Malik") ("segment,S", po::bool_switch(&segment), "segment the image with Perona-Malik beforehand") ("grayscale,g", po::bool_switch(&grayscale), "read in as grayscale") ("video,V", po::bool_switch(&write_video), "enable video output (changes the extension to '.avi')") ("overlay-text,O", po::bool_switch(&overlay_text), "add overlay text") ("invert-selection,I", po::bool_switch(&invert), "invert selected region (see: select)") ("select,s", po::bool_switch(&object_selection), "separate the region encolosed by the contour (adds suffix '_selection')") ("rectangle,R", po::bool_switch(&rectangle_contour), "select rectangular contour interactively") ("circle,C", po::bool_switch(&circle_contour), "select circular contour interactively") ; po::variables_map vm; po::store(po::command_line_parser(argc, argv).options(desc).run(), vm); po::notify(vm); if(vm.count("help")) { std::cout << desc << "\n"; return EXIT_SUCCESS; } if(! vm.count("input")) msg_exit("Error: you have to specify input file name!"); else if(vm.count("input") && ! boost::filesystem::exists(input_filename)) msg_exit("Error: file \"" + input_filename + "\" does not exists!"); if(vm.count("dt") && dt <= 0) msg_exit("Cannot have negative or zero timestep: " + std::to_string(dt) + "."); if(vm.count("mu") && mu < 0) msg_exit("Length penalty parameter cannot be negative: " + std::to_string(mu) + "."); if(vm.count("lambda1")) { if(grayscale && lambda1.size() != 1) msg_exit("Too many lambda1 values for a grayscale image."); else if(! grayscale && lambda1.size() != 3) msg_exit("Number of lambda1 values must be 3 for a colored input image."); else if(grayscale && lambda1[0] < 0) msg_exit("The value of lambda1 cannot be negative."); else if(! grayscale && (lambda1[0] < 0 || lambda1[1] < 0 || lambda1[2] < 0)) msg_exit("Any value of lambda1 cannot be negative."); } else if(! vm.count("lambda1")) { if(grayscale) lambda1 = {1}; else lambda1 = {1, 1, 1}; } if(vm.count("lambda2")) { if(grayscale && lambda2.size() != 1) msg_exit("Too many lambda2 values for a grayscale image."); else if(! grayscale && lambda2.size() != 3) msg_exit("Number of lambda2 values must be 3 for a colored input image."); else if(grayscale && lambda2[0] < 0) msg_exit("The value of lambda2 cannot be negative."); else if(! grayscale && (lambda2[0] < 0 || lambda2[1] < 0 || lambda2[2] < 0)) msg_exit("Any value of lambda2 cannot be negative."); } else if(! vm.count("lambda2")) { if(grayscale) lambda2 = {1}; else lambda2 = {1, 1, 1}; } if(vm.count("eps") && eps < 0) msg_exit("Cannot have negative smoothing parameter: " + std::to_string(eps) + "."); if(vm.count("tol") && tol < 0) msg_exit("Cannot have negative tolerance: " + std::to_string(tol) + "."); if(vm.count("overlay-pos")) { if (boost::iequals(text_position, "TL")) pos = ChanVese::TextPosition::TopLeft; else if(boost::iequals(text_position, "BL")) pos = ChanVese::TextPosition::BottomLeft; else if(boost::iequals(text_position, "TR")) pos = ChanVese::TextPosition::TopRight; else if(boost::iequals(text_position, "BR")) pos = ChanVese::TextPosition::BottomRight; else msg_exit("Invalid text position requested.\n"\ "Correct values are: TL -- top left\n"\ " BL -- bottom left\n"\ " TR -- top right\n"\ " BR -- bottom right"\ ); } if(vm.count("line-color")) { if (boost::iequals(line_color_str, "red")) contour_color = ChanVese::Colors::red; else if(boost::iequals(line_color_str, "green")) contour_color = ChanVese::Colors::green; else if(boost::iequals(line_color_str, "blue")) contour_color = ChanVese::Colors::blue; else if(boost::iequals(line_color_str, "black")) contour_color = ChanVese::Colors::black; else if(boost::iequals(line_color_str, "white")) contour_color = ChanVese::Colors::white; else if(boost::iequals(line_color_str, "magenta")) contour_color = ChanVese::Colors::magenta; else if(boost::iequals(line_color_str, "yellow")) contour_color = ChanVese::Colors::yellow; else if(boost::iequals(line_color_str, "cyan")) contour_color = ChanVese::Colors::cyan; else msg_exit("Invalid contour color requested.\n"\ "Correct values are: red, green, blue, black, white, magenta, yellow, cyan."); } if(vm.count("laplacian-coef") && (L > 0.25 || L < 0)) msg_exit("The Laplacian coefficient in Perona-Malik segmentation must be between 0 and 0.25."); if(vm.count("segment-time") && (T < L)) msg_exit("The segmentation duration must exceed the value of Laplacian coefficient, " + std::to_string(L) + "."); if(rectangle_contour && circle_contour) msg_exit("Cannot initialize with both rectangular and circular contour"); } catch(std::exception & e) { msg_exit("error: " + std::string(e.what())); } //-- Read the image (grayscale or BGR? RGB? BGR? help) cv::Mat _img; if(grayscale) _img = cv::imread(input_filename, CV_LOAD_IMAGE_GRAYSCALE); else _img = cv::imread(input_filename, CV_LOAD_IMAGE_COLOR); if(! _img.data) msg_exit("Error on opening \"" + input_filename + "\" (probably not an image)!"); //-- Second conversion needed since we want to display a colored contour on a grayscale image cv::Mat img; if(grayscale) cv::cvtColor(_img, img, CV_GRAY2RGB); else img = _img; _img.release(); //-- Determine the constants and define functionals max_steps = max_steps < 0 ? std::numeric_limits<int>::max() : max_steps; const int h = img.rows; const int w = img.cols; const int nof_channels = grayscale ? 1 : img.channels(); const auto heaviside = std::bind(regularized_heaviside, std::placeholders::_1, eps); const auto delta = std::bind(regularized_delta, std::placeholders::_1, eps); //-- Construct the level set cv::Mat u; if(rectangle_contour || circle_contour) { std::unique_ptr<InteractiveData> id; cv::startWindowThread(); cv::namedWindow(WINDOW_TITLE, cv::WINDOW_NORMAL); if (rectangle_contour) id = std::unique_ptr<InteractiveDataRect>(new InteractiveDataRect(&img, contour_color)); else if(circle_contour) id = std::unique_ptr<InteractiveDataCirc>(new InteractiveDataCirc(&img, contour_color)); if(id) cv::setMouseCallback(WINDOW_TITLE, on_mouse, id.get()); cv::imshow(WINDOW_TITLE, img); cv::waitKey(); cv::destroyWindow(WINDOW_TITLE); if(id) { if(! id -> is_ok()) msg_exit("You must specify the contour with non-zero dimensions"); u = id -> get_levelset(h, w); } } else u = levelset_checkerboard(h, w); //-- Set up the video writer (and save the first frame) VideoWriterManager vwm; if(write_video) { vwm = VideoWriterManager(input_filename, img, contour_color, fps, pos, overlay_text); vwm.write_frame(u, overlay_text ? "t = 0" : ""); } //-- Split the channels std::vector<cv::Mat> channels; channels.reserve(nof_channels); cv::split(img, channels); if(grayscale) channels.erase(channels.begin() + 1, channels.end()); //-- Smooth the image with Perona-Malik cv::Mat smoothed_img; if(segment) { smoothed_img = perona_malik(channels, h, w, K, L, T); channels.clear(); cv::split(smoothed_img, channels); cv::imwrite(add_suffix(input_filename, "pm"), smoothed_img); } //-- Find intensity sum and derive the stopping condition cv::Mat intensity_avg = cv::Mat(h, w, CV_64FC1); #pragma omp parallel for num_threads(nof_channels) for(int k = 0; k < nof_channels; ++k) { cv::Mat channel(h, w, intensity_avg.type()); channels[k].convertTo(channel, channel.type()); intensity_avg += channel; } intensity_avg /= nof_channels; double stop_cond = tol * cv::norm(intensity_avg, cv::NORM_L2); intensity_avg.release(); //-- Timestep loop for(int t = 1; t <= max_steps; ++t) { cv::Mat u_diff(cv::Mat::zeros(h, w, CV_64FC1)); //-- Channel loop #pragma omp parallel for num_threads(nof_channels) for(int k = 0; k < nof_channels; ++k) { cv::Mat channel = channels[k]; //-- Find the average regional variances const double c1 = region_variance(channel, u, h, w, ChanVese::Region::Inside, heaviside); const double c2 = region_variance(channel, u, h, w, ChanVese::Region::Outside, heaviside); //-- Calculate the contribution of one channel to the level set const cv::Mat variance_inside = variance_penalty(channel, h, w, c1, lambda1[k]); const cv::Mat variance_outside = variance_penalty(channel, h, w, c2, lambda2[k]); u_diff += -variance_inside + variance_outside; } //-- Calculate the curvature (divergence of normalized gradient) const cv::Mat kappa = curvature(u, h, w); //-- Mash the terms together u_diff = dt * (mu * kappa - nu + u_diff / nof_channels); //-- Run delta function on the level set cv::Mat u_cp = u.clone(); cv::parallel_for_(cv::Range(0, h * w), ParallelPixelFunction(u_cp, w, delta)); //-- Shift the level set cv::multiply(u_diff, u_cp, u_diff); const double u_diff_norm = cv::norm(u_diff, cv::NORM_L2); u += u_diff; //-- Save the frame if(write_video) vwm.write_frame(u, overlay_text ? "t = " + std::to_string(t) : ""); //-- Check if we have achieved the desired precision if(u_diff_norm <= stop_cond) break; } //-- Select the region enclosed by the contour and save it to the disk if(object_selection) cv::imwrite(add_suffix(input_filename, "selection"), separate(img, u, h, w, invert)); return EXIT_SUCCESS; }