Multi-focus Image Fusion Method And Application Based On Multi-scale And Deep Learning

As a research field of image fusion,multi-focus image fusion is mainly used to fuse two or more images with different focal points in the same scene to generate a clear image in the field of vision.Since the depth of field of the camera is limited,only the objects within the depth of field can be clearly photographed.This will cause subsequent image processing algorithms(such as feature extraction,target recognition,etc.)to consume more computer computational power,reduce the accuracy of feature extraction and target tracking,and increase the instability of the system.Multi-focus image fusion technology can solve the problem of limited camera depth of field,so as to improve the utilization of image information and increase the reliability of the system.The technology has been used in microscopic imaging,medical and military applications.Pixel level fusion can obtain the most abundant image information,which is the basis of multi-focus image fusion.Accurate recognition of clear areas in the image and extraction of them is the focus of pixel level fusion research.Traditional pixel level fusion method is difficult to accurately identify clear areas,fusion results may appear artifacts,insufficient extraction of image details and other problems.The main research contents of this paper are as follows:(1)For the possible problems of noise and misregistration in multi-focus image fusion,the adaptive filtering denoising algorithm and the fast Fourier transform-based image registration algorithm are respectively selected as preprocessing algorithms for multi-focus image fusion.Experiments show that these preprocessing algorithms can effectively eliminate noise and register images,increasing the stability of multi-focus image fusion algorithm.(2)A multi-focus image fusion algorithm based on dual-tree complex wavelet and pixel convolution neural networks is proposed.Firstly,the p-CNN model is trained by the generated data set,and the low frequency coefficient fusion decision diagram is generated by the trained p-CNN and fused.The high-frequency coefficients are divided into three steps.The absolute value is taken as the largest,and the final decision diagram of each high-frequency coefficient is obtained through gradual refinement and smooth optimization based on the local variance of edges and guided filtering,and then fused according to the final decision diagram.Experimental results show that the algorithm in this paper subjectively improves the effect of preserving image energy and details and removing edge artifacts.Compared with LP?PCNN,NSCT?PCNN,CSR,p-CNN,SF increases 4.4%,3.8%,4.7%,1.4%,EN increases 6%,5.7%,3.9%,0.1%,FD increases 18%,13%,15%,1.6%.The time is about an order of magnitude higher than that of p-CNN,and it also has good fusion effect for the images after registration and high-frequency denoising.(3)A multi-focus image fusion algorithm based on non-subsampled shear let transform(NSST)and sample-dependent convolutional sparse coding is proposed.Firstly,the original image is decomposed by NSST to generate high frequency coefficients and low frequency coefficients.The sample dependence dictionary is obtained by training the data set,and then the low frequency is fused by using the sample dependence convolution sparse coding.The high frequency coefficient uses the absolute value to get the decision graph,and the low frequency decision graph is used as the guide graph to optimize the high frequency fusion coefficient.Finally,the fusion coefficients are inverse transformed to obtain a clear image.A comparative experiment is carried out on the public multi-focus images and tongue images taken in this paper.Experimental results show that subjectively,the algorithm can well preserve image energy,extract image details and eliminate image artifacts.Compared with SR,ASR,CSR,SF is increased by 19%,1.5%,0.5%,EN is increased by 1.3%,0.1%,0.13%,FD is increased by 18.1%,3.2%,1.5%,compared with the comparison algorithm,the fusion speed of the algorithm is increased by 7.3 times,36 times,5.4 times.The experimental results of mismatching and noise multi-focus image fusion show that the algorithm also has good anti-interference performance and is more robust than other algorithms.