Research On Deconvolutional Neural Networks-Based Image Fusion Algorithms

Image fusion is the integration of multiple different images into an image which contains the superior information of all images.As an indispensable preprocessing for image fusion,image registration is also worth studying.Deep learning has the potential to improve the quality of image fusion and registration.Deconvolutional neural networks-based image fusion algorithms are discussed in this thesis.The specific content is as follows:A multi-focus image fusion algorithm based on the deconvolutional neural networks is discussed.The initialized filters are designed for the deconvolutional neural networks.The optimal cutoff frequencies of the initialized filters are obtained by making average gradient and information entropy of the fused images optimal.The deconvolutional neural networks whose filters have been learned is obtained.The images to be fused are decomposed to obtain the feature maps by the deconvolutional neural networks.After the choose-max fusion of the region energy of the feature maps,the feature maps are convoluted with the filters to reconstruct the fused image.The experimental results show that average gradient and information entropy of the fused images obtained by the discussed algorithm are improved compared with random initialization and wavelet method.A multi-focus image registration algorithm based on convolutional neural network and KAZE features is discussed.The KAZE features are detected in the multi-focus images,and they are input to the learned convolutional neural network to generate feature descriptors to perform image registration.The experimental results show that the discussed algorithm improves the matching accuracy compared with the MSER and BRISK features.A MS and PAN image fusion algorithm based on deconvolutional neural networks is discussed.The initialized filters are designed for the deconvolutional neural networks.The optimal cutoff frequencies of the initialized filters are obtained by making information entropy and edge strength of the fused images optimal.The deconvolutional neural networks whose filters have been learned is obtained.The I component of the MS image and PAN image to be fused are decomposed to obtain the feature maps by the deconvolutional neural networks.After the choose-max fusion of the region energy of the feature maps,the feature maps are convoluted with the filters to reconstruct the I component of the fused image.Then the fused image is synthesized with the H and S components of the MS image.The experimental results show that information entropy and edge strength of the fused images obtained by the discussed algorithm are improved compared with random initialization and wavelet method.A MS and PAN image registration algorithm based on convolutional neural network and KAZE features is discussed.The KAZE features are detected in the MS and PAN images,and they are input to the learned convolutional neural network to generate feature descriptors to perform image registration.The experimental results show that the discussed algorithm improves the matching accuracy compared with the MSER and BRISK features.A CT and MRI image fusion algorithm based on deconvolutional neural networks is discussed.The initialized filters are designed for the deconvolutional neural networks.The optimal cutoff frequencies of the initialized filters are obtained by making information entropy and energy of the fused images optimal.The deconvolutional neural networks whose filters have been learned is obtained.The images to be fused are decomposed to obtain the feature maps by the deconvolutional neural networks.After the choose-max fusion of the region energy of the feature maps,the feature maps are convoluted with the learned filters to reconstruct the fused image.The experimental results show that information entropy and energy of the fused images obtained by the discussed algorithm are improved compared with random initialization and wavelet method.The image fusion algorithms based on deconvolutional neural networks discussed in this thesis can successfully improve fusion quality of multi-focus,MS and PAN,CT and MRI images.At the same time,the discussed image registration based on convolutional neural network and KAZE features improves the matching accuracy of the first two types of images.