Research On Image Fusion Algorithm Based On Deep Learning

Image fusion technology is a branch of image processing,it mainly refers to using correlation algorithm to recombine the image information with different imaging principles or bands into a new image.The sharpness and comprehensibility of the fused images are increased,which are more consistent with human vision and further recognition and detection by subsequent computers.Deep learning network has hierarchical structure,it can be a relatively simple way to express complex functions,which can learn the deep feature representation.Thus,image fusion and deep learning are combined to further extract image features to ensure the integrity and clarity of the later image fusion.The research work of this paper mainly includes:(1)Relevant knowledge of image fusion is studied carefully,we graspe the outstanding points and some shortcomings of the existing methods,what's more,discuss and compare several popular image fusion algorithms.Furthermore,lead in deep learning and introduce its development history and current progress,compared deep learning with shallow structure in order to highlight the advantages of depth structure.Finally,summarize the common model methods of deep learning,and introduce the principle and training process of auto-encoder,sparse self-encoder and convolution neural network in detail.(2)A sparse auto encoder model of deep learning to study the image features is selected.Firstly,several original images are segmented by sliding window technique,and then are combined to form a joint matrix.We use a sparse auto-encoder model of deep learning to train the weight matrix and parameter matrix,and fine tune the parameters through the feedback.We get the corresponding features of several original images,these features are automatically learned via the sparse auto encoder,and then the method of maximization selection is used to fuse the original image.According to the uncertainty of the number of neurons,the fusion effect is the best when the number of neurons in the hidden layer is more than the number of input units.The differences between other fusion methods and this method are also discussed.(3)A multi-focus image fusion method of deep auto-encoder with convolution neural network is proposed.It would lose some details such as edge and so on if we regard images as features directly when fusing.In order to avoid this risk,the separation method of the original image is used to separate the focused and the unfocused part of the background and foreground of the source image.Through the auto-encoder,the convolution neural network is trained step by step to complete the construction of the network model.An absolute maximum method is used to fuse the output features of the foreground focused and unfocused part.The output features of background are fused by weighted average method.The feasibility and validity of the algorithm are verified by experiments.