Theoretical Study Of Fusion Method And Its Application In Image Processing

With the development of image denoising algorithm,sparse representation theory and multi source imaging technology,fusion method is becoming more and more important in image processing.There are different denoising performance in different denoising algo-rithms,which often lead to the difference in denoising effect.It is possible to combine these differences to improve the denoising effect.In multi-source images,due to the different imag-ing mechanism,different images carry different information in the same scene.In order to obtain a more accurate and comprehensive image,several source images are often fused by using fusion method.Therefore,the study of effective fusion method in the application of image processing is of great significance and value.In this dissertation,based on frac-tional Fourier transform,image decomposition,sparse representation theory and a neural network method,the applications of fusion methods in image processing are studied.Some new results are given.The main work includes the following aspects:(1)A fusion estimation method based on fractional Fourier transform is proposed.There are three main steps:Firstly,a pre-estimation is made by any two denoising method separately in the spatial domain.Secondly,using these two estimated results as well as their Fourier trans-form,twice Fourier transform and three times Fourier transform,we obtain a fused result in the fractional Fourier transform domain.Thirdly,the inverse fractional Fourier transform is used to obtain a spatial fusion result.Obviously,this state superposition fractional Fourier transform is a fusion method in four different domains,it combines the spatial and frequency domains information skillfully.Experimental results on benchmark test images demonstrate that the proposed method outperforms state-of-the-art stand-alone methods as:BM3D,D-DID,MLP,EPLL and also superior to the fusion methods such as classic wavelet fusion method,PCA fusion method and the state-of-the-art CIEM fusion method in terms of quan-tity value such as the peak signal to noise ratio(PSNR),the structural similarity(SSIM),and visual quality.(2)An image fusion method based on quaternion wavelet transform and sparse repre-sentation theory is proposed.The proposed method complements some defects existing in traditional multi-scale analysis based fusion methods and sparse representation based fu-sion methods.There are three stages:1)The low-pass and high-pass coefficients of each source image are given using quaternion wavelet decomposition.2)The merged low-pass coefficients are obtained using a sparse representation fusion rule,while the fused high-pass coefficients are obtained using the largest absolution values rule.3)The fused image is ob-tained by taking inverse quaternion wavelet transform.The proposed method is compared with the sparse representation,the discrete wavelet transform,the dual-tree complex wavelet transform and the quaternion wavelet transform image fusion schemes in terms of objective qualities and visual effects.Experimental results show that the proposed fusion scheme is effective.(3)In the fusion process of CT and MRI medical images,although traditional multi-scale methods and sparse representation methods have achieved great success,there are still some shortcomings.For example,some hard tissue and soft tissue edges are too smooth using sparse representation method.In order to compensate for these shortcomings,a hybrid fu-sion method based on image decomposition is proposed.There are three main steps:Firstly,the cartoon and texture parts of the source images are obtained using the improved image decomposition method.Secondly,in the large structure cartoon parts,coefficients are fused using the specific cartoon dictionary and the“L1-max norm”principle.In the curve texture parts such as bones and blood vessels,coefficients are fused using the multi-scale nonsub-sampled contourlet transform and the maximum energy rule.Finally,The fused images are obtained by superimposing the fused cartoon parts and the fused texture parts.The experi-ment results show that the proposed method can improve the fusion effect in term of visual effects and objective qualities.(4)Image decomposition,sparse representation and neural networks play important roles in image fusion.For noisy images,fusion and denoising proceed simultaneously,and the learned dictionary is used a universal dictionary or a analytical dictionary in most existing sparse representation based models.In order to further improve the fusion effect,an image fusion method based on cartoon + texture dictionaries combined with neural network denois-ing method is proposed.In the proposed model,denoising and fusion are processed sepa-rately.Instead of a universal dictionary or a analytical dictionary,the tailored cartoon and texture dictionaries are selected.The proposed method is divided into four main steps:1)The denoised results are obtained using external/internal methods for noisy source images.2)The external/internal cartoon and texture dictionaries are learned from the external/internal data,individually.3)The external fusion result based on the external cartoon + texture dictionar-ies sparse representation(E-CTSR)and the internal result based on the internal cartoon+texture dictionaries sparse representation(I-CTSR)are obtained,individually.4)The exter-nal +internal fusion results are obtained by combining E-CTSR and I-CTSR using a neural network method(EI-CTSR).The experiment results show:1)For isomorphic images,such as multi-focus images,EI-CTSR outperforms the stand-alone E-CTSR,I-CTSR,sparse repre-sentation(SR)and adaptive sparse representation(ASR).The divide-and-conquer methods E-CTSR and I-CTSR are also superior to SR and ASR.2)For heterogeneous multi-mode images,E-CTSR outperforms SR and ASR.