Recently, numerous academic papers in the machine learning / computer vision / image processing domains (re)introduce and discuss a “frequency loss

Comparing images in frequency domain. “Spectral loss” – does it make sense? | Bart Wronski

submited by
Style Pass
2021-07-07 03:00:03

Recently, numerous academic papers in the machine learning / computer vision / image processing domains (re)introduce and discuss a “frequency loss function” or “spectral loss” – and while for many it makes sense and nicely improves achieved results, some of them define or use it wrongly.

The basic idea is – instead of comparing pixels of the image, why not compare the images in the frequency domain for a better high frequency preservation?

Unfortunately, current research contains a lot of “try some new idea, produce a few benchmark numbers that seem good, publish, repeat” papers – without going deeper and understanding what’s going on. Beating a benchmark and an intuitive explanation are enough to publish an idea (but most don’t stick around).

I’ve touched upon loss functions in my previous machine learning oriented posts (I’ll highlight the separable filter optimization and generating blue noise through optimization, where in both I discuss some properties of a good loss), but for a fast recap – in machine learning, loss function is a “cost” that the optimization process tries to minimize.

Leave a Comment