Restauração de imagens subaquáticas utilizando modelos de difusão condicional

Abstract

Computer vision in underwater environments faces significant challenges due to image degradation caused by light absorption and scattering. Such degradation reduces contrast, alters colors and compromises the effectiveness of visual perception systems in tasks such as underwater inspection and monitoring. To address this problem, this work developed a conditional diffusion model for underwater image restoration, based on a U-Net architecture with multi-scale attention mechanisms and iterative DDIM sampling. The model was trained and evaluated on the UIEB dataset, using 178 images in the test set, and achieved a mean PSNR of 21.63 dB and SSIM of 0.886. These results represent gains of +4.23 dB in PSNR and +0.113 in SSIM compared with the original degraded images. The results indicate the viability of conditional diffusion models for underwater image restoration, especially in terms of color cast correction and structural detail recovery.