Research On Description And Matching Of Image Local Features

The techniques of image local feature,which has been the focus of research in the field of computer vision.Many tasks in computer vision,such as image classification,image retrieval,3D reconstruction,image stitching and so on,can be regarded as image feature extraction and recognition issues.Most of these tasks are based on the extraction of local features of the image,and the performance of local features plays an important role in these applications.However,images often undergo changes in translation,rotation,illumination,angle of view,and blur.Especially,the images normally contain complex background,noise interference and pose changes in real-world scenes,which brings more challenges to the extraction of local features of the image.Therefore,great efforts are still needed because local features have important theoretical significance and practical value.Based on the analysis of existing technologies,this paper makes an in-depth study on the descriptor and matching of local features.The main research contents and innovations of this paper include:(1)An improved local feature descriptor based on deep learning is proposed.Firstly,this paper analyzed the existing algorithms and pointed out the defects in them.Then,this paper proposes a triple loss which 1)uses cosine similarity instead of L2 distance to compare descriptors and 2)relies on a log form loss function.The network structure and sampling strategy adopted in this paper are introduced in detail.Finally,the proposed algorithm and related comparison algorithm are trained and tested on the Phototour dataset,and the trained model is used to image matching experiments on the Oxford Affine dataset.The experimental result shows that the proposed algorithm achieves excellent performance on both data sets,which proves the effectiveness and superiority of the proposed algorithm.(2)An improved RANSAC algorithm based on grid statistics is proposed.In the local feature matching stage,it is necessary to eliminate a large number of false matches to achieve the purification of the initial matching.Currently,the RANSAC algorithm is commonly used for the purification of matching.After analyzing the shortcomings of RANSAC,based on observation,an improved RANSAC algorithm based on grid statistics is proposed.Different from the traditional RANSAC algorithm which takes all the matching points as input,the proposed algorithm pre-processes the image by meshing.The matching is filtered out to improve the percentage of inliers by counting the matched pairs in the grids and setting a reasonable threshold,which preferably improve the proportion of the inner points in the set.And from the perspective of probability distribution,this paper demonstrates the rationality of this observation.Finally,experiments on standard test images show that the proposed algorithm greatly improves the efficiency,and guarantees the accuracy.