Research On Infrared And Visible Image Registration And Fusion Algorithm

With the rapid development of imaging sensor technology,multi-band image acquisi-tion tasks have obtained high-tech assistance.The disadvantages of single-sensor imaging systems have become increasingly apparent.Multi-sensor imaging systems have ushered in a new era of technology applications.Multi-sensor imaging comprehensively utilizes complementary information and eliminates redundant information,so that the fused image has richer scene content.Infrared and visible are the most widely used images in the field of image processing.Due to their inherent imaging mechanism and natural complemen-tarity,infrared and visible image registration and fusion tasks can clearly and accurately describe scene targets.This paper focuses on the basic theories of infrared and visible im-age registration and fusion,combining the significant advantages of convolutional neural networks,and proposes image registration and fusion algorithms based on convolutional neural networks.At the same time,it addresses the real-time needs of image fusion tasks,Implement model compression on the image fusion network,thereby reducing the size and calculation of the fusion network model.In this paper,the following research is carried out under the technical requirements and background conditions.First,we conduct a detailed study of infrared and visible image registration tasks.Perspective transformation is selected as the experimental image transformation model.Considering the lack of robustness of traditional registration algorithms based on mu-tual information and image features for multi-source image registration tasks,this pa-per proposes a new image registration fusion network framework CADH,which uses the unique advantages of image feature extraction,fully extracting semantic information such as efficient common features in infrared and visible images,and supervised learning algo-rithms for image transformation matrix training and learning to fit the true labels of image datasets,enabling the network model to learn between infrared and visible images Robust registration features improve the robustness of the image registration model.Through a large number of comparative experiments,it is confirmed that the CADH registration network in this paper has efficient image registration performance.Secondly,infrared and visible images fusion is an important modern image enhance-ment technology in the field of computer vision.It aims to effectively combine infrared information and visible light information to generate scene fusion with rich information and stronger expression.image.In this paper,a new image fusion framework based on convolutional neural networks,RDNFuse,is designed for the difficulties of artificially formulating feature fusion rules and inadequate feature extraction based on traditional image fusion algorithms such as multi-resolution.By integrating the advantages of net-work structures such as ResNet and DenseNet,this paper uses The unsupervised algorithm solves the problem of artificially formulated fusion rules,improves the fine extraction of multi-scale features of the image,and the feature reuse and flow in the network framework,thereby improving the final imaging display performance of the fused image.Then,the convolutional neural network in deep learning technology has made rapid development recently,but there are still application restrictions for specific application scenarios such as mobile platforms or hardware platforms with limited computing power.Therefore,this paper designs and develops a series of model compression experiments for the real-time requirements of image fusion tasks.It not only designs a new backbone network structure to reduce network parameters,but also implements model compression strategies such as network model pruning,quantization,and Huffman coding to reduce the fusion network.While reducing the size of the fusion network model,the existing fusion performance of the network model is stably maintained.Finally,through a large number of subjective and objective comparison experiments,this paper confirms that the image registration network framework can efficiently extract robust common features of images,and has a significant performance improvement com-pared to other traditional registration algorithms.For the image fusion task,the image fusion framework designed in this paper improves the network fusion performance by strengthening the feature reuse of the network image.Experimental comparisons with mainstream image fusion algorithms have shown better fusion performance.Finally,by formulating a series of model compression strategies,on the basis of ensuring the fusion performance of the network model,the network model size and calculation volume are greatly reduced,thereby improving the network operation efficiency.At the same time,infrared and visible light color image fusion framework and different resolution image fusion frameworks are designed to further expand the fusion application scenario.