Research on fusion method for infrared and visible images via compressive sensing

In order to obtain a more exact, reliable and better description than a single source image, we need to fuse source images taken from different sensors to a synthetic image. This paper employs infrared and visible images and uses the theory of compressive sensing to study image fusion method. The fusion method based on compressive sensing theory contains three parts: overcomplete dictionary, the algorithm of sparse vector approximation and fusion rule. This paper selects three trained overcomplete dictionaries by K-means Singular Value Decomposition (K-SVD) including the dictionary only using patches from the infrared images, the dictionary only using patches from the visible images and the dictionary using the combined patches, two sparse vector approximations containing orthogonal matching pursuit and polytope faces pursuit algorithms, and two fusion rules covering maximum ℓ1-norm and maximum absolute of entry of sparse vector which is firstly proposed in this paper to study twelve fusion approaches. The experimental results show that the method using orthogonal matching pursuit can provide better fusion results in the condition of the same parameter setting and the same dictionary and fusion rule, and the method using the dictionary only using patches from the infrared images, the fusion rule of maximum absolute of entry of sparse vector and orthogonal matching pursuit takes almost all the largest objective evaluations and the best fusion quality.