PhyCR-UIE: unsupervised underwater image enhancement based on physical imaging model and cross-modal regularization

Aquatic environments cause light to be absorbed and scattered, which often leads to issues in underwater images such as distorted colors, poor contrast, and lost details. Most existing underwater image enhancement (UIE) techniques depend on supervised learning, constrained by labeled image pair availability and quality. Moreover, their loss functions focus mainly on image modality features, lacking cross-modal semantic constraint capacity. To solve these problems, this paper put forward PhyCR-UIE, an unsupervised UIE approach based on a physical imaging model and cross-modal regularization. Three Convolution Neural Network (CNN)-based subnetworks predict overall ambient light, medium transmission map, and scene radiation respectively. The model does unsupervised training under joint homology and reconstruction loss constraints. Additionally, cross-modal regularization is applied to scene radiation via a Contrastive Language-Image Pre-training (CLIP)-based unsupervised loss Numerous tests conducted on actual underwater datasets show our method outperforms representative approaches in both visual appearance and quantitative measurements.