The formation of underwater images is a complex physical process that often suffers from various degradation factors, such as blurriness, low contrast, and color casts, which pose challenges for underwater object detection and recognition tasks. Because of the absence of reference images, learning-based methods that rely on unpaired images have been employed to enhance the underwater images. However, these methods may lose their effectiveness in real-world complex underwater environments. In this paper, we propose a model-driven cycle-consistent generative adversarial network (CycleGAN) model, which is inspired by the underwater image formation model to estimate the background light, transmission map, scene depth, and attenuation coefficient directly. Comprehensive experiments have demonstrated that our approach surpasses the compared underwater image restoration methods in both qualitative and quantitative aspects, providing restored images with satisfactory color saturation and brightness. We also conduct experiments on underwater object detection to illustrate the effectiveness of our CycleGAN in improving the detection accuracy. All our source codes and data are available at https://github.com/Duanlab123/UW-CycleGAN.

Sparse-view and limited-angle computed tomography (CT) are very challenging problems in real applications. Due to the high ill-posedness, both analytical and iterative reconstruction methods may present distortions and artifacts for such incomplete data problems. In this work, we propose a novel reconstruction model to jointly reconstruct a high-quality image and its corresponding high-resolution projection data. The model is built up by deploying regularization on both CT image and projection data, as well as by introducing a novel full sampling condition to fuse information from both domains. Inspired by the success of deep learning methods in imaging, we utilize the convolutional neural networks to embed and learn both the interrelationship between raw data and reconstructed images and prior information such as regularization from the data, which is implemented in an end-to-end training process. Numerical results demonstrate that the proposed approach outperforms both variational and popular learning-based reconstruction methods for the sparse-view and limited-angle CT problems.

Non-line-of-sight (NLOS) imaging aims to reconstruct the three-dimensional hidden scenes by using time-of-flight photon information after multiple diffuse reflections. The under-sampled scanning data can facilitate fast imaging. However, the resulting reconstruction problem becomes a serious ill-posed inverse problem, the solution of which is highly likely to be degraded due to noises and distortions. In this paper, we propose novel NLOS reconstruction models based on curvature regularization, i.e., the object-domain curvature regularization model and the dual (signal and object)-domain curvature regularization model. In what follows, we develop efficient optimization algorithms relying on the alternating direction method of multipliers (ADMM) with the backtracking stepsize rule, for which all solvers can be implemented on GPUs. We evaluate the proposed algorithms on both synthetic and real datasets, which achieve state-of-the-art performance, especially in the compressed sensing setting. Based on GPU computing, our algorithm is the most effective among iterative methods, balancing reconstruction quality and computational time.