Physics-guided underwater image restoration gains come from bias not physics
Physics-Shaped, Not Physical: Auditing Physics-Guided Underwater Image Restoration
Computer Vision and Pattern Recognition
Summary
Underwater photos often look hazy, so special computer programs try to fix them using ideas from physics. This paper looked closely at one such program and found that the physics part doesn’t actually work as expected. Instead, the program’s success comes from the way it’s designed to think like physics, not from using real physical data. This matters because it shows that pretending to use physics isn’t the same as using real physics when cleaning up underwater images.
What this means in practice
- •For underwater imaging engineers: Design underwater image cleanup algorithms knowing that physics-shaped biases, not actual physics, improve performance under limited training data.
- •For machine learning developers: Develop image restoration models that incorporate inductive biases inspired by physics to improve data efficiency without relying on precise physical models.
Authors
Mochamad Rizal Fauzan, Irgi Surya
Abstract
Physics-guided networks for underwater image restoration predict transmission and backscatter and invert the image-formation model, yet rarely test whether this physics does any work. We audit one representative design, PhysGuide-Net, with four controlled experiments: five-seed ablations on UIEB with zero-shot evaluation on the complete LSUI (4279 pairs) and AUIED3K (3000 pairs) sets; a data-efficiency study with a capacity-matched control that removes the formation model; and a validation of the predicted maps against measured range on Sea-thru. With all 800 training images the physics module gives no zero-shot gain (mean PSNR differences of at most 0.16 dB; all 95% intervals include zero) at 3.4x the FLOPs. With 80 images it helps more (+0.65 dB on UIEB, +1.48 dB on AUIED3K; seed-to-seed s.d. on AUIED3K from 1.3 to 0.5 dB), although not individually significant after correction, and the control does not reproduce this (0.8-1.0 dB lower). Yet the predicted transmission does not behave like a physical quantity: its correlation with range is no stronger than for the range map of a different image (pooled difference +0.02, versus -0.06 for the dark-channel prior), and a Beer-Lambert fit explains 0.4% of its variance. The effect of the module also changes sign with degradation severity between datasets. The benefit thus comes from a physics-shaped inductive bias, not from physics. We further show that single-run ablations can reverse sign and document implementation defects that invalidate loss ablations.