Research on underwater image optimization and super-resolution methods based on multiscale feature fusion

Abstract

Underwater images are generally blue-green in colour, have low contrast, are affected by scattering blur, and lack detail at a low resolution. To address the above degradation problems in a unified manner, this paper proposes a multi-scale feature fusion method for underwater image optimisation and super-resolution. The Way is to introduce an underwater imaging degradation model as a physical constraint. Estimate transmission and colour correction first to reduce wavelength selective absorption. Then, a local convolution, dilated convolution and residual branch are used to construct a multi-scale texture representation. Channel attention further re-calibrates cross-scale features, and a sub-pixel reconstruction head generates the high-resolution image. Jointly train using reconstruction, structural, edge and colour-consistency losses. Experiments on public underwater datasets and synthetic degradation samples have shown that the proposed method improves PSNR, SSIM, UIQM and UCIQE, and is stable in the presence of heavy turbidity, low illumination, and four fold super-resolution.