Research On Image Fusion Based On Multiscale Transform And Sparse Representation

Image fusion is to generate a high quality image by integrating the multiple source images of the same scene obtained by sensor.The fused image contains more abundant,accurate and concise information than any of the individual source images,which is available for the following work such as human perception,computer recognition and object detection,etc.Image fusion algorithm is studied based on multiscale transform and sparse representation in this thesis.Firstly,the several popular multiscale image fusion algorithms are introduced,and the medical image fusion algorithm is studied based on shift-invariant shearlet transform domain.Then,in order to overcome the shortcomings of traditional wavelet transform,the shift-invariant dual-tree complex shearlet transform and non-subsampled quaternion shearlet transform are constructed respectively,and then analyzing the properties of them.Finally,through the analysis of the current defects about fusion rule,a novel fusion rule is proposed.By combining the new multiscale transform,image fusion algorithms are proposed based on shift-invariant dual-tree complex shearlet transform domain and based on non-subsampled quaternion shearlet transform domain,respectively.The main work is as follows:1.The development status and the image performance evaluation criteria of image fusion are described.And the related conception and theory are briefly introduced,such as shift-invariant shearlet transform,dual-tree complex wavelet transform,quaternion wavelet,compressive sensing,sparse representation and pulse coupled neural network.2.In order to improve the quality of medical image fusion,the medical image fusion algorithm is proposed based on shift-invariant shearlet transform and compressive sensing.Firstly,the source image is decomposed by the shift-invariant shearlet transform.Then,for the low frequency sub-bands,the spatial frequency and regional energy are employed as the activity level measurement.A low frequency fusion rule combining the activity level measurement and similarity matched degree is presented.For the high frequency sub-bands,considering that the high frequency coefficients have abundant data,the fusion rule that compressive sensing is introduced into the pulse coupled neural network is presented.Finally,the fused image is obtained by performing the inverse shift-invariant shearlet transform.The experimental results show that the proposed approach not only improves the quality of fused images but also the efficiency of the algorithm.3.To overcome these disadvantages of traditional wavelet transform,shift-invariant dual-tree complex shearlet transform is constructed by cascading dual-tree complex wavelet transform and shearlet filter banks.A novel infrared and visible image fusion algorithm is proposed based on shift-invariant dual-tree complex shearlet transform and sparse representation.Firstly,the morphology transform is used for the source images to obtain the enhanced images,and the enhanced images are decomposed by the shift-invariant dual-tree complex shearlet transform.Then,for the low frequency sub-bands,the sub-bands are processed by the sparse representation.A fusion rule combining the novel sum-modified Laplacian of sparse coefficients and sigmoid function is presented.For the high frequency sub-bands,a scheme based on the theory of adaptive dual-channel pulse coupled neural network is presented.Finally,the fused image is obtained by performing the inverse shift-invariant dual-tree complex shearlet transform.The experimental results show that the proposed approach greatly improves the clarity and texture features of the fused image,and outperforms other traditional fusion algorithms in terms of both visual quality and objective evaluation.4.To overcome these disadvantages of traditional wavelet transform,non-subsampled quaternion shearlet transform is constructed by cascading quaternion wavelet transform and shearlet filter banks.A multi-focus image fusion algorithm is proposed based on non-subsampled quaternion shearlet transform domain.Firstly,the multi-focus image is decomposed by non-subsampled quaternion shearlet transform.Then,for the low frequency sub-bands,the sub-bands are processed by the over-complete dictionary,and the fused coefficient is obtained by combining 1l-norm,regional energy and sigmoid function.For the high frequency sub-bands,a scheme by combining the spatial frequency,edge energy and similarity matched degree is presented.Finally,the fused image is obtained by performing the inverse non-subsampled quaternion shearlet transform.The experimental results show that the proposed approach can improve the edge and detailed information,and outperforms other classical fusion algorithms in terms of both visual quality and objective evaluation.