Research On Complex Textural Remote Sensed Image Matching

Image matching is a typical problem and difficult technique in photogrammetry and remote sensing,its purpose is to find the point correspondences,in which the points pairs are with the same physical locations.This technique is not only important in the area of photogrammetry and remote sensing,it is but also widely used in machine learning、 computer vision and medical image processing and so on,it is always the research hotspot in those areas.But as the rapid development of sensors and information science,the category and quantity of them become larger,the textures become more complex,it is a challenging task for matching complex textural images.Therefore,designing corresponding matching algorithms for complex textural image matching is important in applications and theoretical researches.By analysing the complex textures in remote sensed images,we classify it into two different categories by its sources:(1)the homogeneous textures and the repetitive textures in images;(2)the textures in multisource images.For both two complex textural image matching problems,our paper has offered corresponding solutions and supplied some algorithms.our study contents are as follows:(1)Considering that the conventional feature matching algorithms cannot match the images with repetitive and homogeneious textures,our paper has proposed a feature matching algorithm combining geometric constraints and belief propagation technique.The feature matching problem has been transformed into an optimization problem in Markov Random Field.According to the similarity measurement between feature descriptors,the unary energy term is designed,according to the geometric constraints,the binary potential function is designed.In this way,the proposed algorithm has sufficiently exploited the geometric information to reduce the influences on feature descriptors caused by complex textures.At last,the optimization problem is solved by belief propagation,this solution corresponds to the matching result constrained by descriptor similarity measurement and geometric information.The experimental results demonstrate that the proposed algorithm can efficiently match images with homegeneous and repetitive textures and has a better matching performance than that of conventional matching algorithms.(2)There exist non-linear radiometric changes between the textures of multisource images,it makes it difficult to complete matching tasks using conventional feature descriptors.To solve this issue our paper proposed a feature descriptor based on oriented magnitude and feature orientation.The proposed method firstly transforms the image into frequency domain by fourier transform,and then utilizes a log-Gabor filter with 4 scales and 6 orientations to filter the image in the frequency domain,then we can obtain the the histogram of oriented magnitude and the histogram of orientation of the minimum moment of phase congruency,both of them are invariant to the non-linear changes between intensities,thus it can effectively reduce the influences of textures in multisource images on descriptors,this is the theoretical foundation of our method.The histogram of oriented magnitude reflects the distribution along each orientation,the second histogram reflects the orientation of features,the proposed descriptor can be obtained by normalizing and concatenating them.Experimental results demonstrate that the proposed descriptor is invariant to nonlinear radiometric differences,the matching performance is not only superior to the classical methods but also more robust than current state-of-art multimodal image matching algorithms.(3)As for multimodal images,besides the idea of improving the descriptor against the radiometric differences,it is also a reasonable solution which directly utilizes the geometric information in images and avoids the exploitation of textures with non-linear changes.Accoriding to this motivation,our paper proposed a graph matching method for multimodal images based on combination of Monte Carlo tree search and random forest.This method first extracts the graph structure from image pairs,then the feature matching problem is turned into a graph matching problem.The gaussian process regression model is utilized to implicitly represent the state of matching points and the geometric transformation model between point sets,then this graph matching problem can be solved by state space search method.Then the proposed method designs a series of features based on the geometric constraints formed in graph matching,these features can be utilized to train the random forest regressor,this random forest regressor will be further applied in the Monte Carlo tree search,by this way the search tree will optimize the searching paths in the matching state space,and the goal of fast graph matching will be accomplished.Experimental results demonstrate that the proposed method can achieve satisfactory results even graphs are with rotation change,deformation,noise and subgraph cropping,the performance of matching multimodal images is superior to classical methods.