Vision Model based Image Fusion in Nonsubsampled Contourlet Transform Domain

A bio-inspired image fusion algorithm in nonsubsampled contourlet transform (NSCT) domain is proposed. Two biological vision models, the visual attention mechanism (VAM) and pulse coupled neural network (PCNN), which extract image features under different scales, are employed. VAM based saliency matching degree is used in NSCT low-pass subbands fusion. This ensures the fusion results integrate the salient content completely while retaining a high degree of visual consistency with the source images. In NSCT band-pass subband fusion, a PCNN motivated by the NSCT band-pass subbands directional coefficients, is employed to extract the image details. The saliency complementarity of different kinds of source images is tested and analyzed. A new fusion quality evaluation index, visual saliency difference, is proposed to measure the performance in terms of visual consistency of the fusion algorithms. Experiments demonstrate that the proposed algorithm significantly improves the quality of fused images, while the proposed visual consistency index can accurately evaluate the visual consistency of the fusion algorithm.