Retinex Variational Underwater Image Enhancement with Hyper-Laplacian and Edge-Preserving Filtering Priors

To address the issues of low contrast, loss of structures and details, artifacts and noise in degraded underwater images, we propose a novel Retinex variational underwater image enhancement method with hyper-Laplacian and edge-preserving filter priors. Initially, a statistical-based technique is first utilized to correct color casts, followed by converting an underwater image from RGB to HSV color space. According to the Retinex theory, the value channel is decomposed into the illumination and reflectance. In our model, the L1/2-norm is imposed on the first-order gradients of both the illumination and reflectance to preserve structures and details of the underwater image. Moreover, an edge-preserving filtering prior is employed to constrain both the illumination and reflectance, eliminating undesirable artifacts and noise. Besides, the L2-norm is simple and effective to constrain the fidelity term. Next, an alternating direction strategy is implemented to optimize this objective. Finally, the enhanced illumination undergoes gamma correction, and the enhanced reflectance is merged to produce the improved value channel. The final enhanced underwater image is obtained by converting the image from HSV space to RGB. Numerous experiments demonstrates that our method excels in enhancing structures and details, as well as removing artifacts and noise on various underwater images. Furthermore, this approach can be extended to applications, such as dehazing and sandstorm removal, showcasing its versatility and effectiveness.