Multi-focus Image Fusion Based On Deep Learning

Due to the depth of field limitation of optical lenses,an image captured by an imaging device produces a phenomenon in which areas within the depth of field range are clearly displayed,while the remaining areas are blurred.In order to solve this problem,image fusion technology came into being.Image fusion can fuse two or more images with different focus to generate a single image,and the fused image can describe the scene more accurately,making the fused image more conducive to human eye recognition and computer processing.Based on deep learning technology,this paper studies the common multi-focus image fusion in image fusion,aiming to improve the quality of multi-focus image fusion.The main research contents of this paper include the following aspects:(1)An improved multi-focus image fusion method based on Dense Fuse network is proposed.Improvements include: adding a multi-scale feature extraction network to the network to obtain more features on the source image,deepening the densely connected network to enhance the information interaction between layers,and introduce sub-pixel convolution to avoid information loss caused by information compression during image reconstruction,introducing a perceptual loss function to improve the overall performance of the network.Compared with the common multi-focus image fusion methods,the fusion results of the improved method not only achieve good results in visual effects,but also improve in objective evaluation indicators.(2)In view of the lack of datasets during network training for the multi-focus image fusion method based on deep learning.A sufficient number of multifocal images are synthesized on the large-scale object detection dataset MS-COCO,and the proposed network is trained by applying random region Gaussian blur to the images.(3)A multi-focus image fusion method based on attention mechanism is proposed.On the basis of multi-scale feature extraction network and densely connected network,a new attention mechanism is introduced to search for important features of source images from two dimensions of space and channel for image reconstruction,which avoids designing complex fusion strategies.In terms of practical application,the proposed method also conducts fusion experiments on multi-focus image sequences.The experimental results show that the fusion effect of the proposed method is better than other multi-focus image fusion methods,and can effectively eliminate the phenomenon of edge contour loss and artifacts.