An autoencoder compresses 8-bit grayscale photos, where every pixel is a whole number from (black) to (white). The reconstruction loss tells you the mean squared gap between the original and rebuilt pixels, but image quality is usually reported as PSNR (peak signal-to-noise ratio), measured in decibels:
Here MSE is the reconstruction loss, the mean of the squared pixel differences, and MAX is the largest value a pixel can hold in this image format. A higher PSNR means a closer copy.
Here is one patch, before and after a trip through the autoencoder.
Original
| col 1 | col 2 | col 3 | col 4 | |
|---|---|---|---|---|
| row 1 | 52 | 140 | 203 | 88 |
| row 2 | 17 | 230 | 96 | 161 |
Rebuilt
| col 1 | col 2 | col 3 | col 4 | |
|---|---|---|---|---|
| row 1 | 60 | 133 | 195 | 94 |
| row 2 | 26 | 216 | 101 | 153 |
What is the PSNR of the rebuilt patch, in decibels? Round your answer to 2 decimal places.