storm_imagery.cc

Go to the documentation of this file.

#include <gvars3/instances.h>
#include <cvd/image_io.h>
#include <cvd/convolution.h>
#include <TooN/wls.h>
#include <tr1/tuple>

#include "storm_imagery.h"
#include "debug.h"
#include "utility.h"

using namespace CVD;
using namespace TooN;
using namespace GVars3;
using namespace std;
using namespace std::tr1;

/**Load all images from disk and do the initial preprocessing. 

@param names List of filenames to load.
@return preprocessed images.
@ingroup gStormImages
**/
vector<Image<float> > load_and_preprocess_images2(const vector<string>& names)
{
    vector<Image<float> > ims;
    //Load images
    for(unsigned int i=0; i < names.size(); i++)
    {
        Image<float> im = img_load(names[i]);
        ims.push_back(im);

        if(ims.back().size() != ims[0].size())
        {
            cerr << "Error with image " << names[i] << ":  all images must be the same size!\n";
            exit(1);
        }
    }
    double mean, variance;
    tie(mean, variance) = mean_and_variance(ims);
    {
        for(unsigned int i=0; i < ims.size(); i++)
            transform(ims[i].begin(), ims[i].end(), ims[i].begin(), bind2nd(minus<double>(), mean));
        for(unsigned int i=0; i < ims.size(); i++)
            transform(ims[i].begin(), ims[i].end(), ims[i].begin(), bind1st(multiplies<double>(), 1/ sqrt(variance)));
    }
    
    tie(mean, variance) = mean_and_variance(ims);

    cerr << "Rescaled:\n";
    cerr << "mean = " << mean << endl;
    cerr << "std  = " << sqrt(variance) << endl;



    //Normalize...

    //Fit the background model
    ImageRef size = ims[0].size();
    Vector<10> p = Zeros;   
    p[6]=-3;
    p[9]=-4;

    Image<Vector<6> > monomials(size);
    Image<double> polynomial(size);
    for(int yy=0; yy < size.y; yy++)
        for(int xx=0; xx < size.x; xx++)
        {
            double x = xx *2./ size.x -1 ;
            double x2 = x*x;
            double y = yy *2./size.y - 1;
            double y2 = yy;
            monomials[yy][xx] = makeVector(1, x, y, x2, x*y, y2);
        }
                    

    for(int i=0;i < 100;i++)
    {
        for(int yy=0; yy < size.y; yy++)
            for(int xx=0; xx < size.x; xx++)
                polynomial[yy][xx] = monomials[yy][xx] *  p.slice<0,6>();

        WLS<10> wls;    
        for(unsigned int i=0; i < ims.size(); i++)
            for(int yy=0; yy < size.y; yy++)
                for(int xx=0; xx < size.x; xx++)
                {
                    double t = i *1. / ims.size();
                    double func = polynomial[yy][xx] * (exp(p[6]*t) + p[8]*exp(p[9]*t)) + p[7];

                    Vector<10> mJ;

                    mJ.slice<0,6>() = exp(p[6]*t)* monomials[yy][xx];
                    //mJ.slice<3,3>() = Zeros;
                    mJ[6] = polynomial[yy][xx] * exp(p[6]*t) * t;
                    //mJ[6] = func  * t;
                    mJ[7] = 1;

                    mJ[8] = polynomial[yy][xx] * exp(p[9]*t);
                    mJ[9] = polynomial[yy][xx] * exp(p[9]*t) * t * p[8];

                    double err = ims[i][yy][xx] - func;

                    double w;

                    
                    if(err > 0)
                        w = .01 / (abs(err) + .01);
                    else
                        w = 1;

                    wls.add_mJ(func - ims[i][yy][xx], -mJ, w);
                }
        
        wls.add_prior(10);
        wls.compute();

        p += wls.get_mu();

        cout << p << endl << endl;
    }
    
    for(unsigned int i=0; i < ims.size(); i++)
        for(int yy=0; yy < size.y; yy++)
            for(int xx=0; xx < size.x; xx++)
            {
                double x = xx *2./ size.x -1 ;
                double x2 = x*x;
                double y = yy *2./size.y - 1;
                double y2 = yy;
                double t = i *1. / ims.size();
                Vector<6> f = makeVector(1, x, y, x2, x*y, y2);
                
                double func = f * p.slice<0,6>() * (exp(p[6]*t) + p[8]*exp(p[9]*t)) + p[7];
                ims[i][yy][xx] -= func;
            }

    tie(mean, variance) = mean_and_variance(ims);

    //A sanity check.
    cerr << "The mean should be small compared to std:\n";
    cerr << "mean = " << mean << endl;
    cerr << "std  = " << sqrt(variance) << endl;

    //Scale by the variance.
    {
        for(unsigned int i=0; i < ims.size(); i++)
            transform(ims[i].begin(), ims[i].end(), ims[i].begin(), bind1st(multiplies<double>(), 1/ sqrt(variance)));
    }
    tie(mean, variance) = mean_and_variance(ims);

    cerr << "Rescaled:\n";
    cerr << "mean = " << mean << endl;
    cerr << "std  = " << sqrt(variance) << endl;

    return ims;
}




/**Load all images from disk and do the initial preprocessing. Currently
this is a high pass filter to make the resultimg images zero mean.

The filter is controlled with the \c preprocess.lpf and \c preprocess.skip Gvars

See also load_and_preprocess_image()

@param names List of filenames to load.
@return preprocessed images.
@ingroup gStormImages
**/
vector<Image<float> > load_and_preprocess_images(const vector<string>& names)
{
    vector<Image<float> > ims;

    //float wide = GV3::get<float>("preprocess.lpf", 0., -1);
    //bool p = GV3::get<bool>("preprocess.skip", 0, -1);
    
    for(unsigned int i=0; i < names.size(); i++)
    {
        Image<float> im = img_load(names[i]);
    
        ims.push_back(preprocess_image(im));
    
        if(ims.back().size() != ims[0].size())
        {
            cerr << "Error with image " << names[i] << ":  all images must be the same size!\n";
            exit(1);
        }
    }
    return ims;
}

/**Compute the mean and variance of the (on average) darkest pixels, in order
to find the correct scaling, by examining hte background.
*/
pair<double, double> auto_fixed_scaling(const vector<Image<float> >& ims, double frac)
{
    assert_same_size(ims);
    
    //Compute the mean image (ish)
    Image<double> ave(ims[0].size());
    ave.fill(0);
    for(unsigned int i=0; i < ims.size(); i++)
        for(int y=0; y < ave.size().y; y++)
            for(int x=0; x < ave.size().x; x++)
                ave[y][x] += ims[i][y][x];
    
    //Find the smallest N% of the pixels...
    vector<pair<double, ImageRef> > pixels;
    for(int y=0; y < ave.size().y; y++)
        for(int x=0; x < ave.size().x; x++)
            pixels.push_back(make_pair(ave[y][x], ImageRef(x,y)));

    int npix = (int) floor(frac *pixels.size() + 0.5);
    npix = max(0, min(npix, (int) pixels.size()));

    nth_element(pixels.begin(), pixels.begin() + npix, pixels.end());

    pixels.resize(npix);
    
    //Now compute the mean and variance of those pixels.
    double sum=0, sum2=0;

    for(unsigned int i=0; i < ims.size(); i++)
    {   
        for(unsigned int j=0; j < pixels.size(); j++)
        {
            sum += ims[i][pixels[j].second];
            sum2 += sq(ims[i][pixels[j].second]);
        }
    }

    double num = 1.0 * pixels.size() * ims.size();
    double mean = sum / num;
    double std  = sqrt(((sum2/num) - mean*mean) * num / (num-1));

    cout << "Automatic determination of fixed scaling:" << endl
         << "mean       = " << mean << endl
         << "std        = " << std << endl
         << "sqrt(mean) = " << sqrt(mean*255)/255 << endl;
    
    return make_pair(mean, std);
}

/**Wrapper for load_and_preprocess_images() to allow more flexible behaviour.

@param files List of filenames to load.
@return preprocessed images.
@ingroup gStormImages
**/
vector<Image<float> > load_and_normalize_images(const vector<string>& files)
{
    //Load the raw data, and then load the spot parameters.
    vector<Image<float> > ims = load_and_preprocess_images(files);
    double mean, variance;
    tie(mean, variance) = mean_and_variance(ims);

    if(GV3::get<bool>("preprocess.fixed_scaling", 0, FATAL_IF_NOT_DEFINED))
    {
        bool skip = GV3::get<bool>("preprocess.skip");
        if(!skip)
        {
            cerr << "WARNING WARNING WARNING WARNING!!!!!!!!!!!!!!!\n";
            cerr << "preprocessing and fixed scaling selected!!!\n";
            exit(1);
        }

        double sub, div;
        if(GV3::get<bool>("preprocess.fixed_scaling.auto", 0, FATAL_IF_NOT_DEFINED))
        {
            double frac = GV3::get<double>("preprocess.fixed_scaling.auto.proportion", 0, FATAL_IF_NOT_DEFINED);
            tie(sub, div) = auto_fixed_scaling(ims, frac);
        }
        else
        {
            sub = GV3::get<double>("preprocess.fixed_scaling.subtract", 0, FATAL_IF_NOT_DEFINED);
            div = GV3::get<double>("preprocess.fixed_scaling.divide", 0, FATAL_IF_NOT_DEFINED);
        }

        for(unsigned int i=0; i < ims.size(); i++)
            for(Image<float>::iterator j=ims[i].begin(); j != ims[i].end(); j++)
                *j = (*j - sub)/div;
    }
    else
    {
        //A sanity check.
        cerr << "The mean should be small compared to std:\n";
        cerr << "mean = " << mean << endl;
        cerr << "std  = " << sqrt(variance) << endl;

        //Scale by the variance.
        {
            for(unsigned int i=0; i < ims.size(); i++)
                transform(ims[i].begin(), ims[i].end(), ims[i].begin(), bind1st(multiplies<double>(), 1/ sqrt(variance)));
        }
    }

    tie(mean, variance) = mean_and_variance(ims);

    //A sanity check.
    cerr << "Rescaled:\n";
    cerr << "mean = " << mean << endl;
    cerr << "std  = " << sqrt(variance) << endl;

    return ims;
}