Image fusion based on guided filter and online robust dictionary learning

Abstract

It has been confirmed that sparse representation (SR) is successfully applied in many fields, including multi-modal image fusion. A novel SR-based image fusion framework is proposed in this paper, which exhibits state-of-the-art performance in not only fusion effects but also computationally efficient. For SR-based image fusion methods, the critical factor is the over-complete dictionary, which makes the input image sparse. A jointly clustered patch online dictionary learning (JCPORDL) method is proposed to construct a lightweight but practical dictionary and also has an advantage in processing large-scale and dynamic data. The clustering of the joint patches helps reduce the amount of training data for the proposed online robust dictionary learning (ORDL) algorithm. Besides, considering the edge-preserving, the guided filter is embedded in the proposed framework. It has right near edge behaviors and will not add much computing burden. In order to verify how the proposed framework superiority, several conventional image fusion methods were used as a comparison. The experiment results indicate that the proposed framework has better effects and more timesaving than SR-based methods with other dictionary learning strategies. Besides, it also has superior performance than mainstream Multi-Scale Transform (MST) based algorithms and Multi-Scale Transform-Sparse Representation (MST-SR) combined algorithms.