PSNR - Peak Signal-to-Noise Ratio

[minhduy13ly]

Metric's name
PSNR - Peak Signal-to-Noise Ratio

Metric's category
Computer Vision

Metrics formula

$$ \text{PSNR} = 10 \cdot \log_{10}\left(\frac{MAX^2}{MSE}\right) $$

where

• MAX is the maximum possible pixel value of the image (e.g for an 8-bit image, MAX = 255).
• MSE is the Mean Squared Error between the original image $$O$$ and the reconstructed image $$R$$. The detailed formula for MSE is (assume both original image and reconstructed image have the size heigh and width as $$h$$ and $$w$$.

$$ MSE = \frac{1}{h \cdot w} \sum_{i=1}^{h} \sum_{j=1}^{w} \left( O(i,j) - R(i,j) \right)^2 $$

Describe the metrics use cases, and any relevant references.
In the field of image generation and reconstruction, we frequently need to measure how closely our generated images resemble the reference images. Besides modern metrics like FID and Inception Score, PSNR and SSIM are two widely-used classical metrics often used for benchmarking.

The intuition behind PSNR is that it expresses how large the possible signal value (MAX) is in comparison to the error (MSE). If the error is subtle compared to the maximum signal (i.e., a high PSNR), the reconstructed image is relatively similar to the target in terms of pixel values.

Additional context
A sample of using PSNR combined with FID for benchmarking could be found in many computer vision papers, such as EdgeConnect: Generative Image Inpainting with Adversarial Edge Learning

[santiviquez]

This is great! Thanks for suggesting it. I'll link it into #69 as metrics to include in the book.

@minhduy13ly Would you like to take the lead on this? I can support you at any step of the PR.

[minhduy13ly]

Sure @santiviquez - I could get help on this and also could help support/review on other metrics in classical ML/computer vision based on my knowledge. Since I have not contribute to any open-source project before and my experience is limited, I would love to receive the feedbacks for further improving.

Do you have any suggestion on the next step?

[santiviquez]

Hi @minhduy13ly, the usual workflow looks like this:

1. Fork this repository
2. Clone the forked repo to your local env
3. Create a new branch before you make any
4. Make the desired changes
5. Commit and push your changes to your forked repo
6. Open a Pull Request from github.com

And that is pretty much it. This video summarizes it nicely: https://www.youtube.com/watch?v=8lGpZkjnkt4

You might need to install latex on your own machine. In case you don't have it, check out this instructions.

Let me know if you have any issues.