How to calcualte psnr of images of two different size?

I am having one image 0f 500 * 500 size. I have resized it to 250 * 250. How to calculate psnr between two images?

Answer 1

[Image Analyst]

Call

p = psnr(image1, image2);

If they're different sizes, then resize one of them.

image2Resized = imresize(image2, size(image1), 'nearest');
p = psnr(image1, image2Resized );

[Ozan Can Sahin]

Dear Image Analyst,

I'm working on image upscaling and if I resize the original image with another interpolation method, calculating PSNR like this wouldn't make sense, or should it make sense? Is there another method to calculate PSNR for two RGB images with different sizes?

[Image Analyst]

It would make sense. If you have another version of the image that had been resized, then you need to resize it to the same size as the original to determine the PSNR to see how much the resizing has degraded it.

[Ozan Can Sahin]

I don't have any other versions of interpolated images or the original image. But does it make sense to downscale the interpolated images that are generated from the original one and then calculate the PSNR of the downscaled one wrt. the original image?

[Image Analyst]

To compute PSNR, you need two images and they must be the same size. You can't compute PSNR on a single image. Perhaps you just want to compute the standard deviation of a single image. PSNR is commonly used to assess degradation of an image that has undergone compression and decompression. So you need the reference image, and the one that has undergone some processing, like a roundtrip compression/decompression.

[Ozan Can Sahin]

I got the idea, and using the function imresize() seems to be the best option, but which method of resizing should I use for the most efficient and proper PSNR value? The same question goes for the SSIM as well. I would use 'bicubic' but I'd like to ask it to an expert first :)

[Image Analyst]

Each method will give a slightly different PSNR. Bicubic should be fine. Use others and compare them if you want.

[Ozan Can Sahin]

I tried both PSNR and SSIM, it worked without errors but I've seen something that doesn't make sense. I've tried comparing Bicubic Interpolation and Nearest Neighbour Interpolation to see if PSNR and SSIM works properly, however I got better results for NN compared to bicubic which doesn't make sense. And this is valid for not only one single image but I've tried it with a few different images. You can see those in the screenshot attached, the code is on the left and the values are on the right. Any ideas why am I getting such results like this?