Research On Registration Methods Of Infrared And Visible Images

In recent years,with the application of various image sensors,the demand for the integration and utilization of image information is increasing.Infrared and visible images are two common spectral images with complementary information.The infrared image can intuitively reflect the thermal radiation energy of the object,and it is not affected by the light,but it will lose the appearance information such as the texture and structure of some objects.While visible images contain rich color and texture information,but are greatly affected by light.In order to make full use of the complementary information contained in the two modal images of infrared and visible,it is necessary to first register the infrared and visible images.The registered images combine the advantages of both infrared and visible images,which is conducive to subsequent image analysis and recognition.Therefore,the research on infrared and visible image registration methods is of great significance.In this thesis registration problem is studied,aiming at the characteristics of infrared and visible spectral images.The specific work is as follows:Firstly,regarding the problem that the resolution and spectral difference between infrared and visible images are too large,this thesis proposes a registration method based on modal transformation.This method designs a second-order network structure.The network first performs a mapping operation on the visible image and transform it into an image with a similar spectrum to the infrared image.Then in the second stage,the network uses the transformed image as a reference,and uses a space transformer and a sampler to perform a transform operation on the infrared image to obtain a registered infrared image which achieves multimodal image registration.The experimental results show that the registration method proposed in this thesis has great advantages in infrared and visible image registration tasks.Secondly,in view of the uncertainty of the mapping relationship between the infrared and visible image registration,this thesis proposes a registration method based on modal transformation and modal independent neighborhood features.Because there are a large number of infrared image areas with the same or similar gray levels,the gray mapping relationship often has non-single correspondence,which causes a large uncertainty in the mapping.For this reason,this thesis considers the image registration method using a combination of modal transformation and modal independent neighborhood features.The visible image is first transformed into an image that approximates the infrared spectrum to reduce spectral differences.Combined with modal independent neighborhood features to represent different types of image features in different modal images,and the ability to make high response to important features.Then the modal independent neighborhood features are extracted based on the transformed image and the infrared image,and the image deformation domain is calculated using the normalized mutual information as a registration metric to obtain a registered image.This method combines the advantages of deep learning automatic extraction features and artificial design features with high response to important features.Experimental results show that compared with the existing registration algorithms,the proposed method has better registration effect.