An efficient and high-quality pansharpening model based on conditional random fields

Abstract

Pansharpening fuses a low spatial resolution multi-spectral (MS) image with the corresponding panchromatic (PAN) image to obtain a high spatial resolution MS (HRMS) image. Traditional fusion methods may easily cause a spectral or spatial distortion when injecting details into an MS image. To preserve the spectral and spatial information, an efficient pansharpening model based on conditional random fields (CRFs) is proposed. With this model, a state feature function is designed to force the HRMS image filtered using a blur function to be consistent with the up-sampled MS image and retain the spectral fidelity. To obtain a proper blur function, a new filter-acquisition method is proposed for the unified CRF-based model. Meanwhile, a transition feature function is defined to enable the transition of HRMS pixels to follow the gradient of a PAN image and ensure the sharpness of the fused image. Considering the characteristics of the gradient domain, a total variation regularization is designed to make the gradient of the HRMS image sparse. Finally, the augmented Lagrangian function of the model is solved by employing an alternating direction method of the multipliers. Experiment results indicate that, compared with previous state-of-the-art pansharpening methods, the proposed method can achieve the best fusion results with high computational efficiency.