I have been trying to determine why is PSNR used instead of SNR for image quality metrics?

I have read through several books and most just say that PSNR is used instead of SNR, but don't give a reason as to why.

I also read these posts:

Signal quality metric


Difference between SNR and PSNR

But I still feel like they don't explain clearly why PSNR is used instead of SNR.

I know that PSNR measures how the peak signal intensity compares to the mean squared error whereas SNR measures signal power compared to noise, but this doesn't help me reach an intuitive conclusion as to why to use PSNR instead of SNR.

Can someone please explain this?

  • $\begingroup$ The PSNR is independent of the particular image content. It can even be computed in the absence of a signal. You don't need to worry about computing a signal amplitude, you just have to know the number of quantization bits. $\endgroup$
    – user7657
    Nov 2, 2015 at 11:10

2 Answers 2


Consider 8-bit quantization. This gives a mean squared error of about 0.577 least significant bit (LSB). Quantizing two images of different average intensities gives two different SNR, because that metric is normalized against the average intensity of the input image, but identical PSNR, because it is normalized against the maximum pixel value of 255. When describing the quality of something like a compression method, it is better to use a metric that is (to an extent) independent of the source material, because it carries meaning even when communicated without providing the input image.

  • $\begingroup$ So I am fine if I compute SSIM and PSNR on raw data also? I noticed many examples quantize the data first, but there are so many to choose from, I thought it would be best to not quantize my data for a fair comparison. $\endgroup$
    – Veridian
    Nov 5, 2015 at 20:14
  • $\begingroup$ I used quantization just as an example of processing that degrades signal quality in a way that is well described by PSNR: addition of quantization noise. $\endgroup$ Nov 6, 2015 at 1:45
  • $\begingroup$ so does the calculation make sense if it is on raw data? I know it can be done, but is it practical and meaningful? $\endgroup$
    – Veridian
    Nov 6, 2015 at 6:07
  • $\begingroup$ What is generating the noise in your case, and how do you measure it? $\endgroup$ Nov 6, 2015 at 7:36
  • $\begingroup$ Noise is generated through different algorithms I'm using, some produce better images than others. I have a gold standard I am comparing against using PSNR and SSIM. The values of the raw data range between 0 and 80. It seems more fair to compute PSNR and SSIM given a peakval of 80 and a dynamic range of 80, than to arbitrarily use an 8-bit jpeg quantization. However, I don't know if the general image-processing community would agree. $\endgroup$
    – Veridian
    Nov 6, 2015 at 17:31

While for most 1D signal people use RMSE / SNR images are different.
While people can't have a feeling about RADAR or Ultra Sound signal they can clearly see whether image quality is good or is it similar to a reference image.

In images it is clear that optimizing image towards RMSE (Or the Squared Error) gives the "Blurred" signal advantage.
Which means edges are blurred.
Our Visual System is very sensitive to edges.

Hence a different metric had to be created.
Someone suggested the PSNR and it became the standard because it matched the visual system better then the RMSE / SNR (It's not good by itself, only better).

Actually, today the topic of a metric for image restoration quality is very hot.
Usually the most accepted metric today is the SSIM (Structural Similarity).

Summarizing it, it is all about being more consistent with the quality our Visual System sees and the RMSE doesn't match it well.

  • $\begingroup$ Could you give a source regarding your sentence: "Someone suggested the PSNR and it became the standard because it matched the visual system better then the RMSE"? I do not understand how is it possible since the PSNR is just a normalized MSE, and that it is not based on the HVS. Thanks! $\endgroup$
    – benlaug
    Dec 15, 2018 at 13:49
  • 1
    $\begingroup$ PSNR is the ratio between the Max Value Squared to the MSE. The only improvement it suggested over MSE is taking into account the DR of values in the images (Which can be different for different format of images). As you wrote it is not HVS based. $\endgroup$
    – Royi
    Dec 15, 2018 at 14:51
  • $\begingroup$ Okay, I understand better what you meant. Thank you for your clarification. $\endgroup$
    – benlaug
    Dec 15, 2018 at 15:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.