Research And Application On Multiscale Image Fusion Algorithm Based On Statistical Model

Multiscale image fusion is one of the mainstream methods of image fusion.In image fusion algorithm,the fusion result is required to include as many details as possible.And it commonly involved saliency and matching measures of regions to be fused,which is often the determinant of fusion effect for the influence of the selection of source details and fusion weights.Commonly,these parameters were defined by statistical methods of multiscale coefficients,while it is not representative enough of the features of regions to be fused.Meanwhile,source images often corrupted by noise,which will lead to the drop of the quality of fusion image and the inconvenience for subsequent observations.To further strengthen the ability of extraction source details and maintaining the fusion effect in noisy environment,this research optimized the definition of saliency and matching measure to improve the fusion result and proposed a new fusion algorithm for noisy source images from the perspective of statistical model.And the specific research work contain the following three aspects.To incorporating more dependency of coefficients in fusion algorithm,a marginal distribution of wavelet coefficients modeled by generalized Gaussian distribution and a joint distribution of wavelet coefficients to be fused modeled by bivariate Laplacian distribution.And the fitting results are pretty good.To improve the fusion effect of current algorithm,this research proposed to incorporate the dependency of coefficients to be fused into fusion algorithm.The matching measure is defined by mutual information and information entropy of marginal and joint distribution;the saliency measure is defined by the asymmetry of kullback-leibler divergence of marginal distribution.Compared with advanced fusion algorithm,the fusion result of medical images,infrared and visible image and multi-focus images in proposed algorithm perform best.The results of this research maintain the most details of source images and great visual effect.The results of objective and subjective evaluation are also excel ent.To effectively suppress the influence of noise for the fusion results,an shrinkage based on anisotropic bivariate shrinkage function is extracted and multiplies by fusion weights,which will achieve noise-reduction effect;and then this research changes the parameter estimation method into moment-based estimation method,which can keep fusion algorithm from noise corruption.Noise-robust fusion algorithm can be achieved by above two steps.And through the contrast experiment,it is found that proposed algorithm perform best in detail preservation and noise reduction.