The absorption and scattering of light in water usually result in the degradation of underwater image quality, such as color distortion and low contrast. Additionally, the performance of acquisition devices may limit the spatial resolution of underwater images, resulting in the loss of image details. Efficient modeling of long-range dependency is essential for understanding the global structure and local context of underwater images to enhance and restore details, which is a challenging task. In this paper, we propose an agent-guided non-local attention network using a multi-color space for underwater image enhancement and super-resolution. Specifically, local features with different receptive fields are first extracted simultaneously in the RGB, Lab, and HSI color spaces of underwater images. Then, the designed agent-guided non-local attention module with high expressiveness and lower computational complexity is utilized to model long-range dependency. Subsequently, the results from the multi-color space are adaptively fused with learned weights, and finally, the reconstruction block composed of deconvolution and the designed non-local attention module is used to output enhanced and super-resolution images. Experiments on multiple datasets demonstrated that our method significantly improves the visual perception of degraded underwater images and efficiently reconstructs missing details, and objective evaluations confirmed the superiority of our method over other state-of-the-art methods.

Underwater images commonly experience degradation caused by light absorption and scattering in water. Developing lightweight and efficient neural networks to restore degraded images is challenging because of the difficulty in obtaining high-quality paired images and the delicate trade-off between model performance and computational demands. To provide a lightweight and efficient solution for restoring images in terms of color, structure, texture details, etc., enabling the underwater image restoration task to be applied in real-world scenes, we propose an unsupervised lightweight multi-branch context network. Specifically, we design two lightweight multi-branch context subnetworks that enable multiple receptive field feature extraction and long-range dependency modeling to estimate scene radiance and transmission maps. Gaussian blur is adopted to approximate the global background light on the twice-downsampled degraded image. We design a comprehensive loss function that incorporates multiple components, including self-supervised consistency loss and reconstruction loss, to train the network using degraded images in an unsupervised learning manner. Experiments on several underwater image datasets demonstrate that our approach realizes good performance with very few model parameters (0.12 M), and is even comparable to state-of-the-art methods (up to 149 M) in color correction and contrast restoration.