Large-scale Building Image Matching

With the development of technology and the popularization of the network,network data has shown explosive growth,and digital life is closely related to people.Among them,largescale building matching is usually regarded as a large-scale image retrieval problem,which can be used for navigation technology and landmark recognition.The research goal of this thesis is to search the target image quickly and correctly in the massive image data,so as to improve the accuracy and efficiency of image retrieval.This thesis has conducted research on image feature learning and retrieval algorithms.The main research work includes:A feature detection algorithm based on deep learning network is proposed.Based on the residual network,the feature weight layer is used to filter out local feature descriptors that represent strong image capabilities,and the model is trained with a public data set.Compared with the traditional local feature extraction method,the comparison experiments on actual buildings and the data of the feature evaluation indicators on the public data set prove that the feature detection algorithm proposed in this thesis can better characterize images and have a higher degree of discrimination when the target is in a complex environment.A feature matching algorithm based on local spatial consistency is proposed.In image matching,in addition to considering the similarity of the feature descriptors,the property of the local spatial consistency of the feature descriptors is also considered.The voting mechanism is used to filter the feature point pairs matched by the nearest neighbor search algorithm,and the correct matching feature point pairs are purified.The comparison experiment using the traditional matching algorithm and the matching algorithm in this thesis verifies the superiority of the matching performance of the matching algorithm proposed in this thesis.An algorithm for large-scale building image retrieval is proposed.Firstly,use the trained strong distinctive feature descriptors based on the residual network and improve the construction of visual vocabulary of these feature descriptors.Secondly,use the Net VLAD algorithm to encode these image feature descriptors to form the features of the building images for retrieval and then use the inverted product quantization algorithm to establish the image index to search for the target image.Finally,use the extended query,that is,the feature matching algorithm based on local spatial consistency is used to verify the retrieval results.By conducting comparative experiments on different image retrieval algorithms on Oxford105 k and Paris106 k datasets,it is verified that the retrieval algorithm in this thesis improves the accuracy and speed of image retrieval compared with other descriptors that characterize images.