Red-channel-degradation-model-for-underwater-image-using-a-variant-of-dark-channel-prior

Red channel degradation model for underwater image using a variant of dark channel prior

The importance of numeral quality measuring of enhancements made by any restoring method is crucial since image enhancement is mostly certified by visual inspection. Like any other areas, perfect images are not available in many situations. In case of underwater images, we are highly posed this way. The dataset provided in [1] contains 890 underwater images in which a few non-degraded images can be found. These perfect images are best choices to opt for reference inputs. Therefore, we requirea mathematical model for red channel degradation. The problem of measuring the amount of improvement after enhancing image is crucial provided that we form a degraded image from an perfect image that exhibits no haze and blurring. So you can add haze to your haze free (reference) images and then apply your restoring method to estimate the haze free image back. In this way, you can measure the quality of your restoration method by computing mean square error or L2 Norm and etc. The following code provides a model based on medium transmission in underwater images. The medium transmission is computed based on Dark Channel Prior. To estimate medium transmission, it is supposed that a haze free environment is colourful with some shadows. Therefore, at least one colour channel has some pixel with low intensity. This assumption is called Dark Channel Prior. Here, we are going to employ a variant of dark channel prior [2] suitable for hazing underwater images. The equation that is used to calculate medium transmission is as follows:

The following degradation model is used on the red channel:

I(x) = J(x)t(x) - A ( 1 – t(x) )

Where J(x) is the original image (not degraded) and I(x) is the degraded image. To increase the amount of degradation, the term A ( 1 – t(x) ) is subtracted from J(x)t(x). To show an example of degrading red channel using the above model for an almost clear underwater image see the image below.

Red Channel Degraded image - Original Image

[1] https://li-chongyi.github.io/proj_benchmark.html

[2] Automatic Red-Channel Underwater Image Restoration

Code: Use this script to run the mediumtransmissionMat() function. Assign the path to your image into fileNameDataSet. Notice: the parameter gsdestruction is set to 3 but feel free to change it with any greater odd number.

      im=imread(fileNameDataSet);
      im2 = im;
      %%%%%% destructing image
      gsdestruction = 3;
      [medtransMat , globalBackgLight]  =  mediumtransmissionMat ( im , gsdestruction ) ;
      im2( : , : , 1) = im2uint8( im2double( im( : , : , 1) ) .* medtransMat - ( globalBackgLight(1) ) * ( 1 - medtransMat) );

Definition of mediumtransmissionMat:

    function [medtransMat , globalBackgLight]  =  mediumtransmissionMat ( im , gs ) 
    %%%%%% gs must be an odd num
    im = im2double(im);
    imheight = size ( im , 1 ) ;
    imwidth = size ( im , 2 ) ;
    %% find brightest pixel in the dark channel- red
    maxvalred  =  max  ( max ( im (  :  ,  :  ,  1  )   )   ) ;
    indxmaxvalred = find ( im ( : , : , 1 )  == maxvalred ) ;
    %% green global background light- scalar
    greenglobalBackLight = im ( indxmaxvalred ( 1 ) +imheight*imwidth ) ;
    %% blue global background light- scalar
    blueglobalBackLight = im ( indxmaxvalred ( 1 ) +2*imheight*imwidth ) ;
    globalBackgLight = double ( [maxvalred , greenglobalBackLight , blueglobalBackLight] );
    %%  medium transmission size is as im
    mediumtransmission = zeros ( imheight , imwidth ) ;
    %%  gs is grid size
    for r = 1:imheight
        for c = 1:imwidth
    %%      four corners of rect
          rmin = r- ( gs-1 ) /2;
          rmax = r+ ( gs-1 ) /2;
          cmin = c- ( gs-1 ) /2;
          cmax = c+ ( gs-1 ) /2;
    %%      if rmin is out of boundary
          if  ( rmin<1 ) 
            rmin = 1;
            rmax = rmin + gs - 1;
          endif
    %%      if cmin is out of boundary
          if  ( cmin<1 ) 
            cmin = 1;
            cmax = cmin+gs - 1;
          endif
    %%      if rmax is out of boundary
          if  ( rmax>imheight ) 
            rmax = imheight;
            rmin = rmax - gs - 1;
          endif
    %%      if cmax is out of boundary
          if  ( cmax>imwidth ) 
            cmax = imwidth;
            cmin = cmax - gs - 1;
          endif    
    %%    get gs by gs patch from image
          patchImage =  im ( rmin:rmax , cmin:cmax , 1:3 );
    %%    find the min of each patch- scalar
          minpatchred = min ( min ( 1 - patchImage ( : , : , 1 ) ) );
          minpatchgreen = min ( min ( patchImage ( : , : , 2 )  )  ) ;
          minpatchblue = min ( min ( patchImage ( : , : , 3 )  )  ) ;
    %%    normalize green and blue by local back light
          patchImage ( : , : , 1 ) = patchImage ( : , : , 1 ) / ( 1 - maxvalred );
          patchImage ( : , : , 2 )  = patchImage ( : , : , 2 ) /greenglobalBackLight;
          patchImage ( : , : , 3 )  = patchImage ( : , : , 3 ) /blueglobalBackLight;
    %%    min of green and blue patches-scalar
          minpatch = min ( minpatchgreen , minpatchblue ) ;
          minpatch = min ( minpatch , minpatchred );      
    %%    medium transmission- scalar
          mediumtransmission ( r , c )  = 1 - minpatch;     
        endfor % for c    
    endfor  % for r
    medtransMat = mediumtransmission;
    end  % end of function