Hyperspectral Dehazing Using Admm-Adam Theory

ADMM-Adam theory, originally invented for hyperspectral remote sensing, is adopted to solve the challenging ill-posed hyperspectral dehazing problem for the first time, aiming to recover lost spatial information in hazy hyperspectral images covered by haze. Under this innovative framework, benefitted from combining deep learning (DL) and convex optimization (CO), we can train the neural network using just small data and still obtain promising results. In this work, we implement a simple Unet model to obtain a rough DL solution, contributing to the complexity reduction of network parameters, besides avoiding the time-consuming training procedure. Furthermore, experimental evidences show that plugging existing peer methods into the powerful ADMM-Adam framework also yields improved dehazing performance. To conclude, we demonstrate the feasibility and flexibility of solving complicated inverse problems (e.g., hyperspectral dehazing) with a dramatic enhancement under the ADMM-Adam framework.