Research On Multi-Source Images Fusion Algorithm And Optimization Acceleration Technology

In modern society,we are facing a large amount of data from different sources,which are often used separately.Combining multi-source images data can provide more useful information,such as infrared and visible images fusion.The fusion algorithm can make the fused image content more accurate and comprehensive,with richer information,and more suitable for observing,analyzing,and understanding images.Therefore,multi-source images fusion has become a hot issue in the field of image processing.Image fusion has broad development prospects in fields such as machine vision,biological research,military,and security.In order to pursue concise and efficient fusion methods,as well as apply deep learning methods for higher quality image fusion,research on multi-source image fusion algorithms is carried out.However,in engineering applications,it is limited by computational resources and requires real-time processing.Therefore,it is necessary to optimize and accelerate the research on fusion algorithms to meet the practical needs of engineering applications.On the basis of studying the multi-source image fusion algorithm based on guided filter,this article proposes an infrared and visible images fusion algorithm based on anisotropic diffusion and fast guided filter.The algorithm first extracts the detail layer and base layer of infrared image and visible images through anisotropic diffusion algorithm,then obtains the fusion weights of detail layer and base layer respectively through the side window Gaussian filter and the fast guided filter,and finally reconstructs the fused base layer and detail layer to obtain the final fusion image through the additive fusion strategy.On the TNO dataset,the fusion effects of algorithms such as ADF,GFF,TIF,GTF were analyzed and compared.The results showed that this algorithm improved by 5.2%,2.2%,1.8%and 0.8%on STD,AG,SF and EN indicators compared to the suboptimal algorithm,and optimized by 4.3%on N^AB/F artifact indicators.We optimized and improved the Swin Fusion image fusion algorithm based on Swin Transformer to address the shortcomings of image detail and texture information loss.We proposed an infrared and visible images fusion algorithm based on cross domain fusion and norm optimization.Norm optimization is used in the fusion network to generate the pre-fused image,and Conv Ne Xt network is used to extract the depth features of the pre-fused image,supplement the detailed texture information,and improve the loss function of the network to reduce the artifact information.On the TNO dataset,this algorithm was analyzed and compared with Fusion GAN,Dence Fuse,PIAFusion,PMGI,RFN-Nest and Swin Fusion algorithms.The results showed that this algorithm improved by 2.4%,6.0%,2.6%and 0.9%on MS-SSIM,AG,SF and CC metrics and optimized by 26.3%on N^AB/F artifact metrics.In addition,the ablation experiments of pre-fusion images and loss function are carried out for the designed algorithm,and the results show that the improvement of the designed algorithm is effective.To meet the real-time and engineering requirements of the fusion algorithm,the key steps of the infrared and visible images fusion algorithm based on anisotropic diffusion and fast guided filter are optimized and accelerated using the pipeline design and high-speed parallel computing characteristics of FPGA.The logical implementation of the algorithm is completed using Verilog-HDL.Comparing the simulation results of Model Sim and Matlab,it is shown that the optimized and accelerated algorithm has achieved an optimization improvement in computational time,which can be reduced from 1559.42ms to 19.48ms,and can meet the real-time requirements on the TNO dataset.