Research On Multi-source Image Matching And Fusion Methods For LiDAR And Visible Images

Multi-source image matching and fusion is at the base in multi-source information processing,it is a research hotspot and has important research value.LiDAR images can record the distance and intensity information,and visible images can obtain the color and spectral information.Therefore,the matching and fusion of visible and LiDAR images are widely used in military and civilian tasks,such as precise guidance,scene classification,and target recognition.Due to the differences in grayscale distribution and texture style,geometric deformation and noise interference in multi-source images,the matching and fusion of visible and LiDAR images is a challenging task.This dissertation conducted a series of studies: first,we study LiDAR image denosing to improve image quality and matching performance;second,we study visible and LiDAR image matching to spatially align multi-source images;finally,we study visible and LiDAR image fusion to combine the complementary information.The main research contents and innovations include:In the research of LiDAR image preprocessing technology,an intensity image denoising method based on adaptive non-local means(NLM)and wavelet threshold shrinkage is proposed.This method uses homomorphic transformation to convert multiplicative noise into additive noise and uses wavelet transform to decompose the image into low frequency and high frequency coefficients.Adaptive NLM and threshold shrinkage are used for low frequency and high frequency coefficient denoising,respectively.Experiments show that this method can effectively suppress speckle noise in LiDAR intensity images while protecting the image details.Taking advantage of the similarity of the local structural features of multi-source images,a new local feature descriptor,histogram of oriented structure maps(HOSM),is proposed.This method first extracts the multi-directional structure maps of the image,and then uses the guided filtering to enhance the consistency of the structure maps in multi-source images.Good matching performance shows that HOSM is robust to the non-linear changes between visible and LiDAR intensity images.Inspired by the principle of spectral imaging and image gray-scale mapping theory,an image matching method based on regularization generation adversarial network is proposed.This method first designs a new generative adversarial network for image transformation,and then uses local features to establish multi-source image matching relationships.Experiments show that image transformation can effectively eliminate the differences between visible and LiDAR intensity images and improve the feature matching performance.Inspired by the directional,local and band-passing characteristics of the biological vision system,a pixel-level visible and LiDAR intensity image fusion method based on total variation and saliency analysis is proposed.This method decomposes the image into a base layer and a detail layer.The base layer fusion uses total variation,and the detail layer fusion uses weighted averaging.The weight map is based on saliency analysis.Experiments show that this method maintains complementary information in multi-source images,and the fused image is more suited for visual perception and machine analysis.To synthesize the spectral information of the visible image and the spatial information of the LiDAR range image,a feature-level multi-source image fusion method based on total variation and extreme morphological profiles(EMP)is proposed.First,we propose a new EMP feature to extract structural features from multi-source images.Then,the corresponding feature maps in the EMPs of the visible and LiDAR range images are fused by total variation to obtain a fused EMP.Experiments show that feature fusion can improve the scene classification accuracy,because EMP can effectively represent the complementary features of visible and LiDAR range images,and the total variation model can effectively fuse the complementary information.