Research On Image Retrieval Method In 3D Reconstruction

Large-scale 3D reconstruction is a core problem in computer vision and digital photogrammetry.It is a process of recovering the camera's momentary location and obtaining scene three-dimensional information only through a set of images with overlapping degrees.How to quickly and accurately obtain images with overlapping regions in large-scale image data is a key issue in 3D reconstruction.In the traditional aerial triangulation process,strict camera calibration parameters and regular navigation band information are needed to obtain similar information between images.In the field of computer vision,the Content-Based Image Retrieval(CBIR)is usually used to obtain similar images.This method does not require any information outside the image,and the retrieval speed is fast,manual intervention is less,and the degree of automation is higher in the process of processing.In view of the current mobile data,digital camera data,Internet data and unmanned aerial images and aerial images that can't or can hardly be obtained,the content based image retrieval method can accurately and quickly retrieve similar images.Therefore,it is necessary to study how to use computer vision to solve the problems in digital photogrammetry.The basic idea of the content based image retrieval method is to express the image by extracting the visual features of the image,and then index a large number of visual features,so as to quickly and accurately retrieve the similar visual features,and finally get the images with the same or similar content.Therefore,the main work of CBIR focuses on the extraction of visual features and the construction of the index of mass characteristics,in which the extraction of visual features mainly revolves around the texture,color,shape and spatial geometry of the image,and the index structure is mainly a tree type index structure and a hash based index structure.In the background of large scale image retrieval,how to quickly and accurately retrieve feature data with high data and vector dimension has become a hot and difficult point in the current research work of CBIR.Based on the above content,the contents of this paper include:1.This paper summarizes the basic theory and algorithm principle of the content based image retrieval method,sums up the flow of the content based image retrieval method,and summarizes the two key problems of the image retrieval in the 3D reconstruction.One is how to extract the visual content accurately in the face of the content of the aerial images.The feature descriptors are invariable to the scale,rotation,translation and illumination changes of the image.Two is how to establish the index structure of the extracted mass and high dimensional feature descriptors,so as to quickly and accurately obtain the similar features,analyze the descriptor data,and the different index structure will lead to the different retrieval results.2.In this paper,the principles and implementation details of two characteristic operators of SIFT and SURF are studied.The LIFT feature operator based on depth learning is introduced.The advantages and disadvantages of the three feature extraction algorithms are analyzed,and the stability and robustness of the three operators under different image visual information are compared through a large number of experiments.The SIFT feature extraction is slower,but the robustness is high.The SURF operator is an improved algorithm of the SIFT operator.Its extraction speed is faster and its robustness is higher for some types of images.The LIFT operator is a feature extraction operator based on depth learning recently proposed.It has a very large ascending space and is a big direction of future research.3.Detailed analysis and implementation of three kinds of index structure of random KD tree,random projection tree,vocabulary tree,random KD tree to accelerate the retrieval of massive data through the establishment of a random forest,random projection tree through the use of certain strategies,high-dimensional data The dimensionality reduction is to avoid the "dimension disaster" caused by highdimensional data,and the vocabulary tree establishes the index relationship for massive data by establishing visual words,which greatly accelerates the speed and precision of data retrieval.In this paper,the performance of the three algorithms and the effect of retrieving large amounts of data are compared in detail.4.By combining the retrieval algorithm with the feature extraction algorithm,the image retrieval process in the large-scale 3D reconstruction is realized quickly and accurately.The image is extracted by SIFT and SURF algorithm respectively,and the tree type index structure is constructed for the obtained mass feature descriptors,and then the feature operator to be retrieved is used.In the tree structure,the similarity descriptor is retrieved.Finally,the image is fed into the image by a certain strategy,and the image has the overlapping region of the image to be retrieved.The results verify the accuracy and practicability of the method.Aiming at different tree index structure,by improving the traditional data similarity measurement strategy,it improves the robustness and accuracy of retrieval to a certain extent.