Answers are grounded in this paper and cite their sources.
Underwater optical imaging is inevitably degraded by particle scattering and wavelength-dependent absorption, which jointly cause image blur, low contrast, and color casts. These optical degradations impose three critical requirements on underwater image enhancement for robotic vision systems: accurate extraction of local details, effective modeling of global scene structure, and real-time processing. However, existing CNN- or transformer-based architectures alone often fail to satisfy all three demands simultaneously. To address this, we propose a progressive transformer-CNN (TCP) model for real-time underwater optical image restoration. TCP integrates a transformer encoder, a CNN-based U-Net, and a frequency-spatial fusion stage with cross-stage feature fusion, and employs a progressive restoration structure. In the reported comparison, TCP obtained the highest PSNR and SSIM among the tested methods on EUVP and UFO-120, with a reported inference time of 0.032 s per 256×256 image on a desktop TITAN RTX GPU. This is equivalent to approximately 31 images per second in that desktop setting. Performance on embedded GPUs or underwater robot platforms was not evaluated.
No takes yet. Share an insight, caveat, or question.
Su et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: