A general fusion framework based on multiscale transform and block compressed sensing sampling for multifocus images

Multiscale transform (MST) and compressed sensing (CS) are the widely used tools in signal processing. This paper proposes a general fusion framework by taking the superiorities of MST and block-based CS for multifocus images. In the framework, the MST is firstly performed on each source images to get their approximate subbands and detail subbands. Then the saliency maps are obtained by taking difference of the approximate subbands and the original source images, followed by the focused region detection using block CS sampling and morphological opening and closing operators, which is used to guide the selection fusion of the approximate subbands and the detail subbands. Finally, the fused image is got by the inverse MST on the merged subbands. In particular, six widely used MSTs, including the shift-invariant and shift-variant transforms, that is, Laplacian pyramid, discrete wavelet transform, stationary wavelet transform, nonsubsampled Laplacian pyramid, nonsubsampled contourlet transform and nonsubsampled shearlet transform, are used as the MST options to test the proposed fusion framework. In the experiments, four pairs of multifocus images are conducted the fusion, and the performance is evaluated by visually effect and objective criteria. The experimental results demonstrate that the fusion framework is effective and superior to the traditional methods.