A Physics-Guided Vision Transformer Framework for Underwater Image Enhancement and Quality Evaluation

Abstract

Underwater images are prone to severe degradation caused by light absorption, scattering and color distortions, which has a great impact on the performance of vision-based marine applications. We propose physics-guided modeling in conjunction with a ViT-based enhancement network to build a hybrid underwater image enhancement framework in this paper. To begin with, a preprocessing module normalizes the input images by resizing, normalizing and slightly augmenting. A physics-based enhancement block removes the attenuation of light in water and the effect of backscatter based on physical model of light underwater image formation. Then, a ViT module is utilized to learn the global contextual information for further enhancement on contrast, color balance and structural information. The outputs of the two modules are then fused to produce the final enhanced image. Performance of proposed method is assessed both from that of reference based and underwater image quality specific indicators with indicators PSNR, SSIM, entropy, sharpness, UCIQE and UIQM. The experimental results demonstrate that the proposed framework can enhance the visual quality and preserve the structural information effectively, which represents great potential for the tasks of underwater object detection and marine exploration.