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

Infrared and visible image fusion algorithms is vital in the domain of image fusion.This fusion technology can be put into use in many territories,for instance military reconnaissance,medical treatment,remote sensing satellites,and other domains.The goal in the infrared image can be survey round-the-clock,but the texture details are not distinct sufficient.In spite of the visible image includes abundant detailed information,because the image has high environmental requirements,it cannot be observed in all weather or bad weather.The fusion of infrared images and visible images is of great significance.Infrared and visible image fusion means are mainly split into two classes,one is based on traditional ways,and the other is based on deep learning ways.The traditional method has an early origin,but it requires manual configuration of fusion rules,high computational complexity,and low fusion efficiency.Therefore,in recent years,deep learning methods can automatically update the weights and can break through the limitations of traditional methods to a certain extent.However,since there is no fusion image for infrared and visible images that can be used as a reference,supervised learning methods cannot be used for fusion,and the existing deep learning fusion methods have problems of brightness degradation or unclear texture information.In this thesis,two unsupervised deep convolutional neural network algorithms are proposed to fuse infrared and visible images.Mainly innovate in two aspects of network structure and objective function.The research contents are as follows:(1)First,an unsupervised network structure with orthogonalization of convolution kernels is proposed to reduce the view redundancy by calculating the distance between the view parameters without orthogonality and the orthogonality as one of the optimization updates of the loss function.Secondly,the gradient loss in each direction is used to ensure that the fusion image is more similar to the visible image,and finally,the brightness weighting method is used to avoid the brightness degradation problem in the fusion result.(2)The infrared and visible image fusion task is treated as a conditional weighted model,thus transforming it into a conditional probabilistic learning task.A pair of conditional probabilities are generated by constructing a multi-level cross-coupling network structure.By extracting different levels of cross-coupled conditional features in the encoder,the decoder can directly generate a pair of conditional probabilities to generate fusion results.An efficient loss function consisting of weighted fidelity and structural loss for LC saliency and brightness trains the network.Compared with the existing other fusion methods,the two algorithms are able to reach better fusion consequence.Compared with the best depth algorithm,the image fusion assessment criteria of the first algorithm enhanced by 10% on mean,while the second algorithm improved by 20% on average.