A deep Retinex network for underwater low-light image enhancement

Underwater images suffer from color cast and low contrast due to the light absorption and scattering. Especially when natural light is not sufficient, large dark areas appear in the captured image, making it impossible to understand the image content. To address this issue, we propose an underwater low-light enhancement method based on Retinex theory. Our model is an end-to-end trainable. The decomposition network decomposes the raw image into reflectance and illumination according to Retinex theory. In the reflectance enhancement network, cross-residual blocks and dense connections can improve the efficiency of feature utilization and the hybrid attention concentrate on the regions of interest in feature maps from different perspectives. The illumination adjustment network utilizes adaptive frequency convolutional blocks to generate additional band information, which reconstructs the more natural illumination. In order to preserve the color consistency of the enhanced image with the reference image, we project the HSV space into the Cartesian coordinate system and use the Euclidean distance as the color cast loss to constrain the enhancement network. Qualitative and quantitative evaluations on different underwater datasets indicate that our method has the excellent performance and can achieve delightful visual enhancements.