Upscaling can't recover what downscaling deleted — I measured how much
Here's an experiment you can reproduce. Take an image with fine structure — thin gridlines, text, sharp edges — shrink it to a quarter of its side length, interpolate it back up to the original size with bicubic, and compare against the original pixel by pixel. I measured two things: PSNR, and high-frequency energy (the mean absolute difference between neighbouring pixels — how much edge and…
In this experiment, an image with fine structure was reduced to a quarter of its size and then enlarged back to its original size using bicubic interpolation. The results were measured using two metrics: PSNR (Peak Signal-to-Noise Ratio) and high-frequency energy. At a 4x enlargement, PSNR was 17.4 dB, and high-frequency energy dropped to 20.5% of the original.
At a 2x enlargement, the results were better, with PSNR around 52% and high-frequency energy at 80%. The reason behind this loss is that the high-frequency information carried by thin lines and sharp edges is lost when the image is reduced in size. Interpolation can smooth edges but cannot recreate the lost information. Super-resolution, on the other hand, approaches the problem differently.
It doesn't interpolate but rather guesses the likely structure based on the vast amount of real photographs it has seen during training. Models like Real-ESRGAN, SwinIR, and Real-HAT-GAN learn the relationship between clean and degraded images, allowing them to fill in the likely structure during the enlargement process. The key takeaway is that super-resolution adds a model's guess, not the detail that was originally present in the image.
While it enhances the visual appearance of the image, it is not suitable for applications like ID photos, forensic analysis, or OCR, where the original information is required.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
