CMRG-CycleGAN: Color-line module and retinex guided CycleGAN for underwater image enhancement

Abstract

The complex underwater environment leads to light attenuation and scattering, resulting in color distortion, haze, blurring, and low-light conditions in images, which hinder underwater operations. Some existing deep learning methods rely too heavily on the quality of training data, and some reference images in the datasets do not conform to the principles of optical imaging; thus, the results often fail to meet expectations. This paper presents an underwater image enhancement algorithm, named CMRG-CycleGAN, which combines a color line model with a Retinex-guided CycleGAN to address these challenges. A multi-scale color line model (MCLM) is designed to endow the enhancement branch with physical modeling capabilities, improving the enhancement quality. Additionally, a reverse Retinex model (RRM) is designed in the degradation branch. A new joint optimization model is employed to process the illumination and reflection components to degrade the intermediate image, resulting in a reconstructed image that aligns more closely with the principles of underwater imaging. Each branch utilizes a twin-network generator to independently encode the detailed subnetwork and the color line subnetwork or Retinex subnetwork, improving the physical modeling capabilities of the network. Finally, the training of the discriminator is constrained using a relative adversarial loss, which further improves the autonomy of the network. Subjective and objective analyses on benchmark datasets demonstrate that the proposed CMRG-CycleGAN achieves strong performance in both visual quality and evaluation metrics.