Research Of Multi-focus Image Fusion Algorithm Based On Convolutional Neural Network

In real life,people hope to get a clear image of all the objects in the same scene when using a camera to shoot.But the camera lens is limited by the depth of field and cannot focus on all the objects at the same time,so some areas in the images are clear and some areas are blurred.Multi-focus image fusion technology can effectively solve this problem by fuse multiple images with different focus areas in the same scene into a full clear image,improving the information utilization rate of the image.Accurate recognition and extraction of the focus region in the source images is a difficult problem in the multi-focus image fusion algorithm.If the focus regions are extracted incompletely,it will lead to the phenomenon of artifacts and loss of edge contours in the fusion results.Based on the analysis of multi-focus image features,this paper proposed a new multi-focus image fusion method.The main contents and results are as follows:(1)In this paper,the advantages of convolutional neural network in image feature extraction is used to train a convolutional neural network model based on Siamese network for multi-focus image fusion.In this method,the generation of the decision map in the process of multi-focus image fusion is regarded as a binary classification problem.Firstly,a convolutional neural network model is trained to distinguish the focusing pixels and blurred pixels in the source image to obtain the initial weight graph;then through pixel interpolation and morphological processing to optimize the weight graph to get a fusion decision map.The experimental results show that this method can improve the accuracy of detection of focused and non-focused regions.(2)In order to further accurately distinguish the boundary of focused and non-focused regions and improve the problem of artifacts in fusion results.a multi-focus image fusion method based on robust principal component analysis and convolutional neural network is proposed.Firstly,the source images are decomposed by robust principal component analysis to obtain the low-rank and sparse components;secondly,the fusion decision map of low-rank components is constructed by convolutional neural network,and the sparse components are fused by Sum-Modified-Laplacian and guided filter.Experimental results show that the proposed method can better preserve the texture details of the source images,avoid the generation of artifacts and improve the quality of the fused images.