Upscaling images: single-output

In this tutorial you will learn how to use the 'dnn_superres' interface to upscale an image via pre-trained neural networks.

Building

When building OpenCV, run the following command to build the 'dnn_superres' module:

cmake -DOPENCV_EXTRA_MODULES_PATH=<opencv_contrib>/modules -Dopencv_dnn_superres=ON <opencv_source_dir>

Or make sure you check the dnn_superres module in the GUI version of CMake: cmake-gui.

Source Code of the sample

// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.

#include <iostream>

#include <opencv2/dnn_superres.hpp>

#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>

using namespace std;
using namespace cv;
using namespace dnn;
using namespace dnn_superres;

int main(int argc, char *argv[])
{
    // Check for valid command line arguments, print usage
    // if insufficient arguments were given.
    if ( argc < 4 ) {
        cout << "usage:   Arg 1: image     | Path to image" << endl;
        cout << "\t Arg 2: algorithm | bilinear, bicubic, edsr, espcn, fsrcnn or lapsrn" << endl;
        cout << "\t Arg 3: scale     | 2, 3 or 4 \n";
        cout << "\t Arg 4: path to model file \n";
        return -1;
    }

    string img_path = string(argv[1]);
    string algorithm = string(argv[2]);
    int scale = atoi(argv[3]);
    string path = "";

    if( argc > 4)
        path = string(argv[4]);

    // Load the image
    Mat img = cv::imread(img_path);
    Mat original_img(img);
    if ( img.empty() )
    {
        std::cerr << "Couldn't load image: " << img << "\n";
        return -2;
    }

    //Make dnn super resolution instance
    DnnSuperResImpl sr;

    Mat img_new;

    if( algorithm == "bilinear" ){
        resize(img, img_new, Size(), scale, scale, 2);
    }
    else if( algorithm == "bicubic" )
    {
        resize(img, img_new, Size(), scale, scale, 3);
    }
    else if( algorithm == "edsr" || algorithm == "espcn" || algorithm == "fsrcnn" || algorithm == "lapsrn" )
    {
        sr.readModel(path);
        sr.setModel(algorithm, scale);
        sr.upsample(img, img_new);
    }
    else{
        std::cerr << "Algorithm not recognized. \n";
    }

    if ( img_new.empty() )
    {
        std::cerr << "Upsampling failed. \n";
        return -3;
    }
    cout << "Upsampling succeeded. \n";

    // Display image
    cv::namedWindow("Initial Image", WINDOW_AUTOSIZE);
    cv::imshow("Initial Image", img_new);
    //cv::imwrite("./saved.jpg", img_new);
    cv::waitKey(0);

    return 0;
}

Explanation

1. Set header and namespaces

   #include <opencv2/dnn_superres.hpp>
   using namespace std;
   using namespace cv;
   using namespace dnn;
   using namespace dnn_superres;

   If you want you can set the namespace like the code above.

2. Create the Dnn Superres object

   DnnSuperResImpl sr;

   This is just to create the object, register the custom dnn layers and get access to the class functions.

3. Read the model

   path = "models/FSRCNN_x2.pb"
   sr.readModel(path);

   This reads the TensorFlow model from the .pb file. Here 'path' is one of the pre-trained Tensorflow models' path file. You can download the models from OpenCV's GitHub, in the 'dnn_superres' module.

4. Set the model

   sr.setModel("fsrcnn", 2);

   Depending on the model you want to run, you have to set the algorithm and upscale factor. This is to know the desired algorithm and scale, even if you change the .pb file's name. For example: if you chose FSRCNN_x2.pb, your algorithm and scale will be 'fsrcnn' and 2, respectively. (Other algorithm options include "edsr", "espcn" and "lapsrn".)

5. Upscale an image

   Mat img = cv::imread(img_path);
   Mat img_new;
   sr.upsample(img, img_new);

   Now we can upscale any image. Load an image via the standard 'imread' function and create a new Mat for the destination image. Then simple upscale. Your upscaled image is located in 'img_new'.

Original:

Upscaled Image via FSRCNN:

Upscaled Image via Bicubic Interpolation: