Research On Infrared And Visible Image Fusion Algorithm Based On Deep Learning

Image plays an important role in human visual perception.Image fusion technology aims to utilize the advantages of different visual signals in the same scene to generate a single image information description and improve the performance of low,medium and advanced visual tasks.Image fusion is a kind of image enhancement technology,aiming to make comprehensively use of the information captured by multi-source sensors,improve the quality of the image and the utilization rate of the image information.The target information detected by infrared image and the rich scene information contained by visible image have become the hot research object of image fusion due to their unique information complementarity,which has a wide range of applications in military,medicine,agriculture and other fields.At present,relying on the powerful feature extraction,representation and reconstruction capabilities of end-to-end learning,deep learning has become the mainstream technology of image fusion research.Compared with traditional image fusion techniques,the performance of image fusion models based on deep learning is significantly improved.With the deepening of deep learning research,some novel theories and methods have also promoted the development of image fusion technology,such as Residual network,Siamese network,Generative confrontation network,Attention mechanism,Structural similarity loss and Perceptual loss function,etc.In view of the problems existing in the fusion methods of infrared and visible images,the following algorithms are proposed in combination with the advantages of traditional image processing technology and deep learning network.Experiments show that the fusion effect of the algorithms proposed in this paper is better.The main research content of this paper is as follows:(1)Infrared and visible image fusion algorithm based on N-RGAN modelAiming at the problems of loss of details and poor contrast in the fusion of infrared and visible light images,we propose an image fusion method based on N-RGAN network.Through the image decomposition advantages of NSST,the image is decomposed on different scales to obtain high-frequency sub-band coefficients and low-frequency sub-band coefficients.In order to retain more detailed information,the trained GAN network is used to complete the fusion of high-frequency sub-bands,and the source infrared image is used as the criterion so that the fusion result can retain more target information.The edge information is used to construct the weight map to guide the fusion of low-frequency sub-bands,so as to improve the image clarity and contrast.Finally,image reconstruction is performed through the inverse transformation of NSST to obtain fusion results.Through the comparison of various algorithms,it shows the feasibility and effectiveness of the proposed algorithm.(2)Infrared and visible image fusion algorithm based on attention twin networkAiming at the problems of poor image fusion effect,high network space-time complexity and poor definition of existing algorithms in complex environments,this paper proposes a self-coding fusion framework based on attention mechanism and twin networks.First,traditional image processing technology is used to scale decomposition of the image to obtain three components of infrared and visible image.For the low frequency component,choosing the improved Laplacian energy and weighted fusion strategy to improve the resolution of the fusion image.For intermediate frequency and high frequency components,twin attention networks are built to extract in-depth features,and attention mechanisms are introduced to improve the network’s processing ability for complex information.By designing fusion strategies,the intermediate frequency and high frequency fusion components are obtained.Finally,the fusion result is obtained by image reconstruction according to the obtained component coefficients.The experimental comparison shows that the proposed algorithm has good fusion effect,simple network structure and strong applicability.