Kernel-driven robust sparse fuzzy subspace clustering and rolling guidance filter for image fusion

Infrared and visible image fusion is often limited by light degradation and thermal radiation diffusion in practical applications; thus, weak contrast and fuzzy structure are usually realized in fusion results. To resolve these, the paper proposes a new infrared and visible image fusion framework based on kernel-driven robust sparse fuzzy subspace clustering and rolling guidance filter algorithms. Firstly, to eliminate the adverse effects of ambient light, a customized homomorphic filter preprocessing was constructed, which cuts off the amplification path of noise while enhancing global contrast, and provides source data. Secondly, a hybrid decomposing approach that joins the nonsubsampled shearlet transform and rolling guidance filter is proposed, which effectively blocks the disordered diffusion of infrared thermal energy and accurately partitions an image into a low-frequency base layer, a low-frequency detail layer, and a high-frequency layer. For heterogeneous components, an asymmetric feature-driven fusion algorithm was designed. For low-frequency components, the kernel-driven robust sparse fuzzy subspace clustering algorithm was innovatively introduced, and spatial consistency and boundary protection were achieved through a manifold-structure constraint of high-dimensional space. On the other hand, a parameter-adaptive-pulse coupled neural network driven by local gradient energy was designed to achieve injection of extremely sharp edges in high-frequency components. Extensive experiments on several benchmark datasets, composed of visual evaluation, objective metrics, and algorithm generalization verification, showed that the proposed algorithm achieved the highest level of both structural fidelity and information richness. Our code is available at https://github.com/0011star/Clustering-Image-Fusion.git .