devEtal.m

% Multi - level semantic segmentation.



addpath('./scripts/')
clc;
[BETA]=calculateBeta;



% =======================================================
% Vectorizing all the images and its corresponding labels
% For training the multi-variate normal distribution.
% =======================================================

all_index = linspace(1,32,32);
my_folder = './HYTA/images';

file_list = dir(fullfile(my_folder, '*.jpg'));
training_images = 25;
testing_images = 7;


% Log file where individual image's performance are recorded

fileID = fopen('./results/devetalresult.txt','a');
fprintf(fileID,'sky \t thin-cloud \t thick-cloud \n');

number_of_exps = 10;

sky_array=zeros(1,testing_images*number_of_exps);
thincloud_array=zeros(1,testing_images*number_of_exps);
thickcloud_array=zeros(1,testing_images*number_of_exps);

index_number = 0;


for exp_number=1:number_of_exps
    
    disp (['Performing experiment number ', num2str(exp_number), ' out of ', num2str(number_of_exps)])

    my_index = randperm(numel(file_list), training_images);
    testing_index = setdiff(all_index, my_index)
    location_dir='./HYTA/images/';

    st_feature_vec=[];
    rows=0; cols=0;
    st_labels=[];


    for kot=1:training_images

        

        FileNames=file_list(my_index(kot)).name;
        I1=imread(strcat(location_dir,FileNames));
        [rows1,cols1,~]=size(I1);

        GroundTruthName=FileNames;
        ind=length(GroundTruthName)-3:1:length(GroundTruthName);
        GroundTruthName(ind)=[];
        GroundTruthName=strcat(GroundTruthName,'_3GT.png');
        GroundTruth=double(imread(['./HYTA/3GT/',GroundTruthName]));
        GT_image1=double(GroundTruth);

        [st_feature_vec1, st_labels1]=features_3labels(I1,GT_image1,BETA);

        st_feature_vec=cat(1,st_feature_vec,st_feature_vec1);
        rows=rows+rows1;
        cols=cols+cols1;
        st_labels=cat(1,st_labels,st_labels1);


    end

    [phi_class1,mu0_class1,mu1_class1,sigma_class1]=likelihood_estimate(st_feature_vec,st_labels,1,2,3);
    [phi_class2,mu0_class2,mu1_class2,sigma_class2]=likelihood_estimate(st_feature_vec,st_labels,2,1,3);
    [phi_class3,mu0_class3,mu1_class3,sigma_class3]=likelihood_estimate(st_feature_vec,st_labels,3,1,2);

    disp ('Parameters for multi-variate normal distribution learnt.');



    % =============================================






    for kot=1:testing_images
        
        index_number = index_number + 1;

        FileNames=file_list(testing_index(kot)).name;    
        I_test=imread(strcat(location_dir,FileNames));        
        [rows_test,cols_test,~]=size(I_test);
        [color_ch_test]=color16_struct(I_test);

        channel0_test=color_ch_test.c1;
        channel1_test=color_ch_test.c5;
        channel2_test=color_ch_test.c13;
        channel0_test=showasImage(channel0_test); channel0_test(channel0_test==0)=1;
        channel1_test=showasImage(channel1_test); channel1_test(channel1_test==0)=1;
        channel2_test=showasImage(channel2_test); channel2_test(channel2_test==0)=1;
        St_zero_test=reshape(channel0_test,rows_test*cols_test,1);
        St_one_test=reshape(channel1_test,rows_test*cols_test,1);
        St_two_test=reshape(channel2_test,rows_test*cols_test,1);
        St_zero_test=St_zero_test./255;
        St_one_test=St_one_test./255;
        St_two_test=St_two_test./255;

        color_test=cat(2,St_one_test,St_two_test,St_zero_test);
        data_vector_test = [ones(rows_test*cols_test,1) color_test]*BETA;
        res_test=reshape(data_vector_test,rows_test,cols_test);
        prob_res_test=(showasImage(res_test))./255 ;
        st_prob_test=reshape(prob_res_test,rows_test*cols_test,1);    

        % This is the feature vector.
        feature_test=st_prob_test;
        label_test=zeros(rows_test*cols_test,1);

        for p=1:rows_test*cols_test

            % This is the log-likelihood estimate for each of the elements of the testing image.   
            likelihood_class1_positivesample= (feature_test(p,:)'-mu1_class1')'*(inv(sigma_class1))*((feature_test(p,:)'-mu1_class1'));
            likelihood_class2_positivesample= (feature_test(p,:)'-mu1_class2')'*(inv(sigma_class2))*((feature_test(p,:)'-mu1_class2'));
            likelihood_class3_positivesample= (feature_test(p,:)'-mu1_class3')'*(inv(sigma_class3))*((feature_test(p,:)'-mu1_class3'));
            like=[likelihood_class1_positivesample,likelihood_class2_positivesample,likelihood_class3_positivesample];
            [~,ind]=min(like);

            label_test(p,1)=ind;

        end

        % Display output
        label_test(label_test==1)=0;
        label_test(label_test==2)=126;
        label_test(label_test==3)=255;

        % Filtering with a 7X7 filter for better results
        A=reshape(label_test,rows_test,cols_test); B = medfilt2(A, [7 7]);


        GroundTruthName=FileNames;
        ind=length(GroundTruthName)-3:1:length(GroundTruthName);
        GroundTruthName(ind)=[];
        GroundTruthName=strcat(GroundTruthName,'_3GT.png');
        GroundTruth=double(imread(['./HYTA/3GT/',GroundTruthName]));
        GroundTruth=double(GroundTruth);
        
        
        % converting them to [128 128] for reporting
        GroundTruth = imresize(GroundTruth,[128 128]);
        GroundTruth(GroundTruth<=63)=0;
        GroundTruth(63<GroundTruth & GroundTruth<190)=126;
        GroundTruth(GroundTruth>=190)=255;
        
        B = imresize(B,[128 128]);
        B(B<=63)=0;
        B(63<B & B<190)=126;
        B(B>=190)=255;


        [sky, thin, thick] = error_score(B,GroundTruth);

        sky_array(1,index_number) = sky;
        thincloud_array(1,index_number) = thin;
        thickcloud_array(1,index_number) = thick;


    end    
    
    
end    




sky_percent = nanmean(100*sky_array) ;
thincloud_percent = nanmean(100*thincloud_array) ;
thickcloud_percent = nanmean(100*thickcloud_array) ;

fprintf(fileID,'%f \t %f \t %f \n', sky_percent,thincloud_percent,thickcloud_percent);


fclose(fileID);



disp (['sky error percentage: ',num2str(sky_percent)]);
disp (['thin cloud error percentage: ',num2str(thincloud_percent)]);
disp (['thick cloud error percentage: ',num2str(thickcloud_percent)]);