Research On Multi Focus Image Fusion Algorithm Based On Deep Learning

In real life,when people take photos with cameras,they always hope that everything in the same scene in the photos is clear,so as to obtain the best imaging effect.However,because the camera lens is limited by the depth of field,the lens cannot focus all targets in the same scene at the same time,which leads to some areas of the photos are clear,some areas are fuzzy,and the imaging appearance is poor.Therefore,researchers began to design algorithms to improve this situation.Multi focus image fusion algorithm is the most common of these algorithms.It can fuse multiple images of different focus areas in the same scene into a clear image of the whole area.The fused full focus image has clear regions,which effectively solves the problem of poor appearance of fuzzy regions in the source image and improves the utilization of image information.In recent years,more and more multi focus image fusion algorithms have been proposed.The common multi focus image fusion methods are easy to lead to color distortion and edge artifacts,which is particularly prominent at the boundary line of focus and defocus.In addition,due to the particularity of the field of multi focus image fusion,the public dataset for training can not be found,which leads to the inability to fully train the network.In order to solve these difficulties,this paper proposes two deep learning networks for multi focus image fusion and a synthetic dataset for training.The main contents are as follows:(1)The first proposed network is a two-stage progressive fusion network,which is composed of the initial fusion block network and the enhanced fusion block network.The network is able to progressively learn the color information and detailed features of the source image through end-to-end mapping.In the fusion process,firstly,the network fuses the color information of the source image through the initial fusion block network to generate the initial fusion image.Then,on the basis of the initial fusion image,the enhanced fusion block network is used to further fuse the detailed features of the source image to generate the final fusion image.(2)The second proposed network is a multi-stage progressive fusion network,which contains three stages.Each stage learns the features of the source image at different scales,and combines the features learned at all scales to generate the final fusion image.In order to make full use of the information of different training stages to improve the stability of the network.The network presents a multi-scale feature aggregation strategy across the stage,and the attention mechanism are introduced to gather the study focus of the network.(3)Mask template and Gaussian filter are used to process the "VOC2012 Dataset3"dataset to generate a synthetic dataset for training,which contains common scenes and objects in real life and can simulate focused images in real life.In addition,qualitative and quantitative evaluation of network fusion results are carried out from subjective and objective aspects.In order to enlarge the difference between contrast images,the concept of difference graph is introduced in subjective evaluation experiment.All the evaluation results show that the proposed two networks have reasonable fusion efficiency and can achieve better fusion effect compared with the most advanced fusion algorithm.