Research On Multi-focus Color Image Fusion Based On Deep Learning

In the past three years,the theoretical research of deep learning(DL)has been a hot topic in the field of multi-focus image fusion.the purpose of multi-focus image fusion is integrating the partially focused images into one single image which is focused everywhere.To achieve this purpose,this paper focuses on deep learning-based multifocus color image fusion methods.The main research results are as follows:(1)Multi-focus image fusion algorithm based on DCNN and QUADTREEThis work mainly solves the problem that the DCNN algorithm does not deal with unknown regions.Firstly,the proposed uses DCNN to solve the source image and restore the image size to obtain a rough focus probability map.Through establishing two autoadapted threshold values,the focus probability map is initially segmented to obtain a trimap.Secondly,it uses QUADTREE to further implement focus detection on the trimap to obtain the refined initial decision map.Then,the small area removal and guided filtering are used to optimize the initial decision map to the final decision map.Finally,the pixel-weighted average rule is used to fuse the source image with the final decision map to obtain the final fused image.Although both the subjective observation and objective data reflect the improvement of the quality of the fusion image,it is found from the decision map that the detailed information is not well extracted.(2)Multi-focus image fusion algorithm based on deep convolutional encoderdecoder networkI This work mainly solves the problem that the existing algorithm cannot extract the detailed information of the focused part.in this paper,by studying the principle and properties of multi-focus image generation,we constructed an effective simulation dataset,designed a powerful deep convolutional encoder-decoder network(DED-Net)and a suitable loss function for training.By using the trained network to solve the source image,we got an alpha value image with detailed information.Finally,the pixel weighted average rule is used to fuse the source image with the alpha value to obtain the final fused image.Although no good results have been obtained on objective data,the main observation is that the decision map can extract detailed information very well and has great research value.