Multi Focus Image Fusion Using Convolution Neural Network

In real life,the effect of taking pictures is sometimes unsatisfactory,so it is necessary to fuse the distinct regions of different focus targets in the same scene to obtain the global clear image,which is called multi focus image fusion.The traditional image fusion algorithm has the problems of complex model design and poor fusion image quality,while the deep learning technology uses neural network for image feature learning,which has fast training and strong functionality.Therefore,this paper uses convolutional neural network to optimize the problems existing in the multi focus image fusion algorithm.The research work of this paper is mainly from the following aspects:(1)Aiming at the multi focus image fusion algorithm,this paper introduces its fusion level,fusion method and fusion process,and studies the application of convolutional neural network technology in image fusion.It compares and analyzes the advantages and disadvantages of various fusion algorithms to lay a foundation for the improvement of subsequent algorithms.(2)Because the traditional image fusion algorithm needs artificial design of local filter,on the basis of analyzing the structure of convolutional neural network and the influencing factors of image fusion,so this paper proposes a multi focus image fusion algorithm based on improved convolutional neural network.In the process of image fusion,it introduces wavelet packet decomposition,and selects different coefficient reconstruction methods according to the energy similarity value of each sub-band.By improving the structure of Siamese network,it designs different layers of network structure and selects different convolution cores to realize multi-scale transformation,At the same time,it introduces extended convolution to reduce the parameters.Using the back propagation training of Adam optimization algorithm,output the probability values of clear image and fuzzy image,and finally get the fusion result graph.The experimental results show that this method improves the index of image fusion and the fusion effect of the algorithm,and enhances the scene adaptability.(3)In order to further strengthen image feature learning and improve image clarity,this paper proposes image fusion algorithm combining unsupervised learning with convolutional neural network.First,preprocess the original image,then use k-means clustering algorithm to cluster the clear / fuzzy features of each image block in the initial stage of convolution neural network,set the same convolution kernel in the network,reduce the amount of parameters,and use different activation functions in different convolution layers;fix the feature vector in the spatial pyramid pooling layer,train to get the focus decision diagram,and use the pixel weighting method to get the fusion image.Compared with the experimental results,the improved algorithm improves the accuracy of the network training,and the image quality is better.