Research On Local Feature Extraction Method Of Image Matching

Image local features are widely used in many fields such as image matching,3D reconstruction,image stitching,and image retrieval.In these tasks,the quality of image local features directly determines the performance of the model.Due to the changes of illumination,view point,blur,noise and so on,the image corresponding to the same object will also vary.Therefore,it is a research difficulty to extract the invariable image local feature in the changeable image.In addition,there are many visually similar objects in natural scenes.Finding different local features in these similarities is also a research difficulty.In summary,it is of great significance to resist all kinds of natural and artificial interferences and to extract superior image local features.This thesis focuses on the image local features of image matching and the main contributions are as follows:Aiming at the research of image local feature extraction,this thesis focuses on two aspects: sample mining and loss function.Firstly,in the aspect of sample mining method,the limitations of the existing sample mining algorithms are analyzed,and a latent hard sample mining algorithm is proposed.This algorithm can mine potential samples,which do not exist in the dataset but conform to the changes in the real scene,from the existing sample information.Secondly,in terms of loss function,we analyze the advantages and disadvantages of several kinds of loss function and the adaptive margin loss function is proposed.This loss function combines local context information and global statistics between the same batch descriptors to adaptively determine the margin,taking full advantage of the sample's personality and commonality.In the meanwhile,the loss function also has a global distance segmentation threshold for dividing the matching and non-matching descriptor pairs.The experimental results on the UBC Photo-tour and HPatches dataset show that both of these way can improve the distinguishability of local feature descriptors and improve the accuracy of image matching.In addition,the article also discussed the impact of training batch size,descriptor dimensions,optimization methods,hyperparameters in the loss function on the learning local descriptor model performance.For the joint learning of image local features,this paper adopts the serial joint learning structure.In addition,a feature point detection model based on multi-level receptive field is proposed in order to achieve better detection of robust and repeatable feature points.Finally,the experimental results on the HPatches and EF dataset show that the joint learning model can achieve better image matching.