Research On The Fusion Algorithm Of Infrared And Visible Light Images

Infrared imaging sensors have the ability to capture thermal radiation from objects,but they are not sensitive to changes in the details of the texture of the scene.The visible light imaging sensor has the ability to capture detailed information of the scene space,but is easily affected by the harsh surrounding environment.Therefore,the fusion of infrared and visible light images can capture the complementary features of each band,display the radiation target and retain the scene information.Infrared visible light image fusion technology has wide application prospects in the fields of intelligence collection,monitoring security and aerospace.In order to improve the fusion quality of infrared and visible light images,this paper mainly studies based on multi-scale transformation,convolutional sparse coding and convolutional neural network technology.The main innovative work is as follows:(1)Since the low-frequency coefficients in the NSCT(Non-subsampled Contourlet Transform)domain do not have sparseness,and the sparseness indicates that the "sliding window" operation destroys the deep spatial structure of the image,the third part of this paper proposes a Convolutional Sparse Representation(CSR)and neighborhood feature infrared and visible light image fusion algorithm,referred to as: NSCT-CSR algorithm.First,the low-frequency subgraph in the NSCT domain is decomposed into the base layer and the detail layer by Gaussian filtering,and then the ADMM(Alternating Direction Method of Multipliers)is used to solve the sparse coefficients to complete the fusion of the characteristic response coefficients of the detail layer.At the same time,a reasonable neighborhood feature is designed according to the clarity measurement function to complete the fusion of high-frequency subgraphs in the NSCT domain.The experimental results show that the NSCT-CSR algorithm effectively extracts deeper and clearer measurement information of the source image,and overcomes the sparse representation of blockiness defects.(2)In view of the problem that the NSCT-CSR algorithm retains the NSCT decomposition and weakens the ability of high-frequency subbands to capture feature information,the fourth part of this paper considers the visual characteristics of the parameter-adaptive PCNN(Pulse Coupled Neural Network)model and proposes a CSR and APCNN(Adaptive Pulse Coupled Neural Network)adaptive infrared and visible light image fusion algorithm,referred to as: CSR-APCNN algorithm.The experimental results show that: the visual effect of CSR-APCNN algorithm is compared with the NSCT-CSR algorithm has been significantly improved.(3)NSCT transformation has direction uncertainty,and convolutional coding requires a given fusion strategy,so the adaptability of low-frequency subbands of NSCT-CSR algorithm and CSR-APCNN algorithm is slightly weak.In response to the above problems,the fifth part of this paper proposes an infrared and visible light image fusion algorithm based on NSST(Non-subsampled Shearlet Transform)and SCNN(Siamese Convolutional Neural Network),referred to as: NSST-SCNN algorithm.The NSST-SCNN algorithm uses twin dual-channel convolutional neural networks to learn the characteristics of low-frequency subbands in the NSST domain to output feature maps that measure the spatial details of the subbands.Then,a measurement function based on local similarity is designed according to the feature map of Gaussian filtering to adaptively adjust the fusion mode of the low-frequency subbands in the NSST domain.The experimental results show that the NSSTSCNN algorithm effectively solves the problem given by the low-frequency subband fusion mode,and at the same time overcomes the defect of manually setting the PCNN parameters.