Research On Key Techniques Of Different-source Sensors Image Fusion

In order to obtain the mulit-faced features of the observation scene and improve the information detection capability of the information detection system and its reliability of the information detection system,image sensors based on different image types(different image sensors)are widely integrated into one detection system.The scene images that gotten by the different sensors have distinct characteristics of information complementation and redundancy.In the field of image fusion technology,the emphasis is to research how to extract complementary information effectively and use the redundant information to present all the complementary information of the different-sensor images as much as possible in a fused image.Based on the above processing,a more accurate and comprehensive description of the scene can be obtained,which facilitates subsequent images processing.The main work of this thesis includes the following aspects:The principles and features of different imaging sensors,the classification of image fusion and the development status are summarized.The principles,the implementation process,the advantages and problems of the existing image fusion algorithms are analyzed.Different fusion algorithms have different advantages and certain disadvantages.This thesis attempts to transform the fusion process of infrared and visible images into an optimization problem,and combines the NSCT-based image fusion algorithm with the structure tensor based image fusion algorithm by the optimization model.The optimization model restrains the gradient of the final fused image and the NSCT decomposition coefficient,making it close to the pre-fusion gradient and the pre-fusion NCST decomposition coefficient.Through such optimization constraint,the feature extraction capability of the fusion method can be improved,and more edge and detail information is retained in the fused image,thereby achieving better fusion effect.The effectiveness of the algorithm is verified by experiments.Convolutional neural network has been adpoted by more and more researchers in the field of image fusion due to its powerful learning ability and feature extraction capabilities.This thesis first analyzes the image fusion algorithms based on convolutional neural network,which are proposed by Liu Yu et al.,and then it points out the problems.The problems of the original algorithms are insufficient ability to extract detailed information and incomplete measurement methods in similarity.To deal with these problems,the multi-scale transformation fusion framework and the new similarity measurement method are adopted.Subsequently,the effectiveness of the improved algorithm is verified by experiments.