An Unsupervised Dehazing Network With Hybrid Prior Constraints for Hyperspectral Image

Haze pollution in hyperspectral images (HSIs) leads to surface information lack and image clarity degradation, which seriously affects the performance of subsequent image interpretation. Existing model-based hyperspectral haze removal methods enjoy good interpretability and generalization, but they can only process images in a specific wavelength range due to the principle limitation. Deep learning-based dehazing methods have good feature extraction capability, but the cost of acquiring sufficient training data is high in practical applications. At the same time, taking into account that HSIs have spectral low-rank structures, fully utilizing the low-rank property will facilitate the reconstruction of HSIs. In order to combine the complementary benefits of deep learning-based and physical model-based approaches, we decide to formulate HSI dehazing reconstruction as an unsupervised deep image prior (DIP) framework. Specifically, we propose an unsupervised dehazing network with hybrid prior constraints (HPC-UDN) for HSI haze removal, which effectively integrates low-rank prior, deep priors, and physical haze prior. First, the low-rank prior of hyperspectral data is characterized by matrix decomposition, where the decomposition factors are learned through two generative networks. Then, multiple spectral groups are divided based on the correlation and complementarity between spectral bands. In order to exchange information between adjacent spectral groups, a novel spectral grouping feature fusion (SGFF) module is designed, which connects neighboring spectral groups to transfer spectral and spatial features. Finally, high-quality HSI is recovered by merging the features extracted from each spectral group. Extensive simulated and real-data experiments certify the effectiveness and robustness of the presented unsupervised approach and potential applications in the GF-5 image dehazing task.