Location Recognition On Visual Feature Matching

Location identification is a key technology for intelligent systems,involving image feature extraction,feature matching and large-scale indexing.It has extremely important application value in intelligent navigation and visual positioning of intelligent systems.It is a research hotspot in the field of computer vision and artificial intelligence.The thesis studies the local features of images in three aspects: feature extraction,feature matching,image indexing.The main work includes:The traditional manual feature points have limited ability to describe the local area of the image.When the image is blurred,the angle of view or the illumination changes drastically,the invariance of the feature points is weakened.Therefore,the thesis analyzes several commonly used local feature extraction methods.Based on the convolutional neural network,this thesis proposes to use the feature selection layer to extract significant local feature points and their descriptors,and use the labeled building dataset to train the model.This method uses image classification model to solve the problem of image feature extraction.Through the quantitative analysis on the public dataset and the contrast experiments of the actual application scenarios,it is verified that this feature extraction method has strong robustness and distinguishability when the illumination and viewing angle change.The commonly used feature matching algorithm causes a large number of mismatches due to only considering the similarity of feature descriptors.This thesis introduced the concept of local area consistency into feature point matching,this feature matching algorithm is more stable.Compared with the traditional matching algorithm,the spatial consistency information of the image is added to help choose the correct matching result.It can be verified by a large number of contrast experiments that this method is superior to the traditional feature matching algorithm in matching accuracy and operation speed.In large-scale image retrieval,the image descriptors formed by the feature aggregation coding technology affects the accuracy of the retrieval when the background of the image is complex.This thesis proposes to combine feature coding techniques with significant deep descriptors to form more accurate building image vectors.The experiment uses the public image retrieval data set(Oxford5k and Paris6k)to verify that the retrieval method of this thesis can achieve high retrieval precision.