How to Crop and get SSIM Value apply on two images

I was struggling to get ssim value with two images of coins, as the placement position of coin is different little bit

like I have two coin images, 1st image 401x401,2nd image 441x441, on comparision

due to positions of images at different positions ssim(img1,img2) is giving incorrect output, I tried cropping but that didnot worked

Thank You

[DGM]

If you were to register the two images as best you could, what value would the error metric retain? How do you propose to separate the influence of the spatial transformations from the differences in the unregistered images? Why would SSIM be a particularly relevant error metric?

I should add that the fact that the data range of the two images varies by a few orders of magnitude means that everything you do to try to put them on the same scale will become another inseparable component of the useless number you get out of ssim().

[shubham kumar gupta]

I also have a image of same dimension but still in different position, I hae a clean image, and a noisy image, I created a denoiser and tested the noisy image over that algorithm, that was my final output, now to compare how much this image is close to clean image I need to compare it using PSNR and ssim, but due to different positions placing of image I am unable to compare

Now I am getting ssim as 5e-5 which is very low due to the improper placement of both the images, If it is put over one another very close then I guess ssim should come around 0.80+ , that's why I'm stuck here

mat file contains both image data

Answer 1

[DGM]

If it's assumed that there is no scale or rotation transformation, then you can go ahead and try to align the images.

load coin_ssim.mat

% pad reference image since object is so close to edges
refpict = padarray(mat2gray(clean_img),[20 20],'replicate','both');
% crop test image down to extract the object alone
badpict = imcrop(mat2gray(img2),[2.5 61.5 357 363]);

% maximize normalized cross-correlation to find offset
szb = size(badpict);
c = normxcorr2(badpict,refpict);
[idxy idxx] = find(c == max(c(:)));
osy = idxy-szb(1);
osx = idxx-szb(2);

% crop the reference pict to the ROI
refpict = refpict(osy:idxy-1,osx:idxx-1);

imshow(imfuse(badpict,refpict,'checkerboard'));

ssim(badpict,refpict) % a number

ans = 0.2529

[shubham kumar gupta]

Thank you

[shubham kumar gupta]

I've heard of this Image Registration - MATLAB & Simulink (mathworks.com) can we improve this further? Thank you

[DGM]

You can try that if you want. I was under the impression that you created the one image from the other. If you didn't apply a scale or rotation transformation in that process, then you shouldn't try finding transformations you know don't apply.

The above only handles translation. I used the Registration Estimator App to get translation+scale using affine2d transformation.

load coin_ssim.mat
load coin_ssim_reg.mat

refpict = mat2gray(clean_img);
badpict = regobj.RegisteredImage;

ssim(badpict,refpict)

ans = 0.2934

That gives you another number. It's slightly higher than with translation alone -- but if you didn't actually apply a scale transformation in creating the image, then this apparent improvement is not meaningful.

Again, the error caused by imperfect registration is inseparable from the error caused by the denoising/filtering process. You should decide which thing you're trying to measure.

[shubham kumar gupta]

thank you

[shubham kumar gupta]

can you help me increase ssim value 😥😥😥

[DGM]

You're not going to get significantly higher (e.g. 80%) SSIM through any amount of extra registration tweaking. As I've said already, there's little meaning to the SSIM when you cannot know the error contribution. If your goal is to measure the quality of your denoising, then you can't isolate that information from the error caused by the transformations required for registration. Neither is that number independent of the accuracy of the registration estimation itself. Overfitting the registration (e.g. trying to remove scale/rotation/skew when none was originally applied) only makes the number less meaningful.

The SSIM you get describes the end-to-end fidelity of the entire transformation -> noising -> denoising -> registration -> transformation process. If that's what you want, then I guess that the number would be appropriate.

I imagine the vast majority of the error has nothing to do with the remaining registration error. It has to do with the fact that the denoised image is so much darker and blurrier than the source. If that's the image you need to measure, then measuring it is what you get. If it needs to yet be adjusted, then maybe you can play with levels to get closer, or you can apply the filtering suggested in prior threads. I recall @yanqi liu posted a good noise masking example.

load coin_ssim.mat
load coin_ssim_reg.mat

refpict = mat2gray(clean_img);
badpict = regobj.RegisteredImage;

% adjust levels
badpict = imadjust(badpict,[0.15 0.55],[0 1],0.30);

imshow([badpict refpict])

ssim(badpict,refpict)

ans = 0.5497

[shubham kumar gupta]

Thanks a lot @DGM

Answer 2

[yanqi liu]

yes，sir，may be make them to same size，and then compare

load coin_ssim.mat
% get location
bw1 = im2bw(mat2gray(clean_img));
bw2 = imclose(im2bw(mat2gray(img2),0.3),strel('disk',9));
bw2 = bwareafilt(bw2,1);
% make same size
[r,c] = find(bw1);
clean_img = clean_img(min(r):max(r),min(c):max(c));
[r,c] = find(bw2);
img2 = img2(min(r):max(r),min(c):max(c));
img2 = imresize(img2, size(clean_img),'bilinear');
ssim(mat2gray(clean_img),mat2gray(img2))

ans = 0.2620

% compare
figure; imshowpair(clean_img,img2);

figure; montage({mat2gray(clean_img),mat2gray(img2)}, 'Size', [1 2], 'BackgroundColor', 'w', 'BorderSize', [2 2]);

[shubham kumar gupta]

thank you

[shubham kumar gupta]

I've heard of this Image Registration - MATLAB & Simulink (mathworks.com) can we improve this further? Thank you

[shubham kumar gupta]

I'm getting this issue

[DGM]

Uncheck "has rotation". I couldn't get it to behave with the extra degree of freedom either. Granted, I didn't spend that much time on it.

[shubham kumar gupta]

Thank you

[shubham kumar gupta]

can you help me increase ssim value 😥😥😥