I've seen this post before and I always pass it up because it's such a bad image.
The image content has been pretty much destroyed. It was once saved by a camera in JPG and then again it was likely transcoded by some program to an even lower-quality JPG. The final encoding is a roughly 40% quality 4:2:0 subsampled pixel salad. Maybe someone who is more familiar with the language could visually identify the characters based on their rough shape and the surrounding context, but I couldn't.
Still, I might as well give it a shot. Let's start by doing a perspective correction and getting rid of the background. That way we don't have to deal with it skewing all the adjustments.
% this JPG is complete trash ~40% Q, 4:2:0 subsampled
% there is no hope of actually making this readable again
inpict = imread('IMG_1549.jpeg');
inpict = im2double(inpict); % work in FP
% these are the coordinates of the box corners
boxm = [75 161; % [x y]
1 1420;
1020 1430;
945 145];
% assert that this is where they're supposed to be
boxf = [1 1; % [x y]
1 1440;
1080 1440;
1080 1];
% do projective transformation
TF = fitgeotrans(boxm,boxf,'projective');
outview = imref2d(size(inpict));
inpict = imwarp(inpict,TF,'fillvalues',0,'outputview',outview);
Create a mask to try to select the blue highlighter marks. Again, this is more or less hopeless, since most of the chroma information (H and S) has been reduced to 20x20px blocks. The best we can do is get a rough mask and feather it.
% create mask roughly selecting the highlighted area
[H,~,~] = imsplit(rgb2hsv(inpict));
mk = H >= 0.367 & H <= 0.784;
% feather the mask outward since region boundaries are garbage
mk = imdilate(mk,ones(3));
mk = imgaussfilt(double(mk),2);
Try to correct for the highlighter marks:
% reduce image to gray, adjust levels/gamma in ROI
gpict = rgb2gray(inpict);
adjpict = imadjust(gpict,[0.1 1],[0 1],0.55);
% composite images to create gray image without highlighter marks
outpict = replacepixels(adjpict,gpict,mk);
Try to flatten the result:
% flatten image using 'grain extract' blending
% use a sum of gaussian filter (this could be simplified)
lf1 = imgaussfilt(outpict,5);
lf2 = imgaussfilt(outpict,20);
lowfreq = imblend(lf1,lf2,1,'average');
outpict = imblend(lowfreq,outpict,1,'grainextract');
% adjust levels/gamma
outpict = imlnc(outpict,'independent','g',0.8,'k',2);
Could I get more contrast out of it? Perhaps, but it's hard for me to judge at which point readability is being compromised (if it's not already a lost cause). For example, instead of performing flattening via a single 'grain extract' blend, we could use multiple techniques and combine them. The 'grain extract' method has weakness in some of the dark regions, whereas other regions tend to cause problems for the division used by imflatfield(). By using both, we can try to avoid the low-contrast problems in the SW corner.
% flatten image using 'grain extract' blending
% use a sum of gaussian filter (this could be simplified)
lf1 = imgaussfilt(outpict,5);
lf2 = imgaussfilt(outpict,20);
lowfreq = imblend(lf1,lf2,1,'average');
flat1 = imblend(lowfreq,outpict,1,'grainextract');
% flatten image using imflatfield() (division)
flat2 = imflatfield(outpict,7);
% stretch and average the two flattened images
outpict = imblend(imadjust(flat1),imadjust(flat2),1,'average');
% adjust levels/gamma
outpict = imlnc(outpict,'independent','g',0.8,'k',2);
... but I don't know if that's really any better.
In this example, I'm using imlnc(), imblend() and replacepixels() from MIMT, though both could be done directly with little effort. I've posted examples on the forum before which show how to do what imlnc() and replacepixels() do. The 'grain extract' blend is simple enough to figure out.
