Discriminative Dictionary Learning-Based Multiple Component Decomposition for Detail-Preserving Noisy Image Fusion

How to effectively preserve the fine-scale details of the image when noises are suppressed is one of the great challenges faced by scholars in the field of noisy image fusion. The traditional noisy image fusion method tends to smooth the fine-scale structures excessively. To overcome the oversmoothing issue, we develop a novel method that can perform fusion, denoising, and preservation of fine structures simultaneously. In this method, the image is modeled as a superposition of coarse structures and fine details. At the same time, a brand new strategy is developed to decompose the input image into coarse and fine components for the further exploitation of afforded discrimination by learned dictionary. Specifically, to preserve the coarse-scale structures and recover the fine details, a novel discriminative dictionary-learning algorithm is proposed to utilize weighted nuclear norm regularization and sparse constraint to characterize coarse structures and fine components, respectively. For image separation, we present a weighted Schatten sparse nuclear norm regularization and integrate it into the separation model to extract the coarse structures. To estimate the fine components submerged in the noise, we propose to exploit the image & x2019;s nonlocal self-similarity and develop gradient-preservation term based on the gradient histogram constraint. Finally, we develop an innovative fusion rule based on the activity level of the recovered patch to construct the fused coding coefficients of different components. Our experiments show that the proposed method has impressive subjective visual quality and objective metric performance.