Research On Multi-view 3D Reconstruction Based On Hyperspectral Image

Building three-dimensional model of ground object is the basis of characterization of ground object characteristics and remote sensing imaging simulation.In these studies,not only high-precision three-dimensional geometric model but also high-precision optical properties of the three-dimensional model are required.The existing methods are:using 3D geometric modeling equipment and spectral measurement equipment respectively to obtain the 3D geometric model of the target and the spectral characteristics of the target material,and then associating the measured material spectrum with the 3D geometric model through the classification of the target material,so as to realize the high-precision modeling of the 3D geometric and spectral characteristics of the target.Based on the cross field of hyperspectral imaging technology and three-dimensional imaging technology,this paper proposes a method of target three-dimensional model reconstruction based on imaging spectral measurement,and realizes the geometric and spectral fine modeling of target three-dimensional model.The main research work of this paper can be divided into the following points:A multi angle reconstruction method of monocular vision based on optimal band combination is proposed as the standard technical route.In order to reduce data redundancy and improve the efficiency of the algorithm,instead of band by band reconstruction,FMFOA algorithm is used to select the optimal band combination from all bands as the input of image scale space.In the follow-up study,the 3D point cloud registration operation which affects the accuracy of the model is reduced to the image matching operation in the two-dimensional space.This dissertation proposes a dichotomous fluctuation fruit fly optimization algorithm(FOA)to select the optimal band combination of hyperspectral images.Although there are many spectral segments in hyperspectral images,which contain abundant spectral information and can represent the physical and chemical properties of ground objects,the number of bands in the three-dimensional modeling stage will cause data redundancy and affect the modeling efficiency,so the optimal band combination is used to reconstruct the image scale space.The traditional FOA algorithm will appear premature phenomenon when solving complex optimization problems,so the convergence mode of the traditional FOA algorithm is improved,and an improved FOA algorithm based on binary fluctuation model is proposed.The performance of the improved algorithm is verified on the benchmark function data set and the benchmark hyperspectral image data set,and the band selection of the self built hyperspectral image data is completed.An F-SIFT feature extraction algorithm based on spectral space is proposed.The image scale space of traditional SIFT algorithm is obtained from the initial image through continuous Gaussian blur and downsampling operation.Although the image space is obtained,most of the data are not completely real,so the number of feature points extracted by the algorithm is small.In this dissertation,the image pyramid is reconstructed by using the same parameters of the hyperspectral images in scale,position and target object.The image pyramid is constructed by using the optimal band combination image as the bottom input of the pyramid and the image pyramid is constructed by continuous sampling.Because the operation of Gauss fuzzy is abandoned in the process of pyramid construction,each layer of image obtained represents the real information of the image at this scale,so the number of feature points of the image increases greatly.The experimental results show that although the reconstructed image pyramid increases the number of feature points,the increasing scale reduces the efficiency of the algorithm.Therefore,the fast eight neighborhood criterion is proposed to determine each pixel of each layer of image in the pyramid.If the probability of the pixel becoming a feature point is lower than the threshold,the candidate feature point set is removed.A dual position constraint criterion is proposed.After the extraction of feature points,the most commonly used feature point matching method is the ratio of nearest neighbor to next nearest neighbor.However,experiments show that the matching result of this method is greatly affected by the ratio,and the error rate increases when the ratio is large.Therefore,a double position constraint criterion is proposed,and the whole matching process is divided into two steps.The first step is still to use the ratio of nearest neighbor to next nearest neighbor for rough matching of feature points.In order to make the final result contain as many correct matching pairs as possible,the ratio threshold is set to above 1.0,and the similarity degree of each pair is recorded in the matching process.The second step is to eliminate the mismatches by double position constraints,and sort the matching results in the previous step according to the degree of similarity.Because there are errors in the matching process due to the setting of threshold,four matching pairs are randomly selected from the 20 matching pairs with the highest degree of similarity as the benchmark matching pairs.The matching is determined according to the relative position relationship between the feature points of the matched pair and the datum match.If the difference between the actual calculated position and the theoretical position is within the threshold range,the matching is considered to be the correct match match,otherwise,the matching is removed from the candidate match.In the stage of spectral mapping after obtaining the 3D model,different mapping methods are selected according to whether the selected feature points belong to the original feature point set or the derived feature point set.Therefore,two kinds of spectral mapping models are designed.For the original feature point,the coordinates of the exit image corresponding to the feature point are found by backtracking,and then the position of the feature point in the next frame is calculated according to the iterative position optimization criterion.The spectral mapping process of derived feature points is divided into three steps,including point cloud attribute classification,optimal plane estimation and dimension reduction coordinate calculation.With the assistance of spatial projection geometry and evolutionary algorithm optimization theory,the spectral mapping model is constructed.