ICL-Net: Inverse Cognitive Learning Network for Remote Sensing Image Dehazing

Abstract

When imaging the Earth's surface, space-based optical imaging sensors are inevitably interfered by scattering media, such as clouds and haze, resulting in serious degradation of remote sensing images they capture. To enhance the quality of remote sensing images and mitigate the influence of clouds, haze, and other media, we construct a novel approach called the inverse cognitive learning network. The network mainly consists of multiscale inverse cognitive learning blocks that we designed. It has the capability to extract image features at multiple scales, adaptively focus on the global information and location-related local information, and effectively constrain the haze. In the multiscale inverse cognitive learning block, we embed the designed inverse cognitive learning module and parallel haze constraint module. The inverse cognitive learning module simulates the inverse process of human brain cognitive image, and gradually learns the haze information from the depth, moderate, and breadth channel features. The parallel haze constraint module integrates the extracted haze information through a dual-branch approach to realize strong constraints on haze features. Experimental results indicate that our approach notably enhances the clarity of remote sensing images that suffer from cloud cover and haze, and possesses more perfect haze removal effect and robustness than state-of-the-art dehazing approaches.