Research On Image Feature Extraction And Matching Based On KAZE Algorithm

As a very important part of image processing,image feature extraction and matching technology have played an important role in many fields such as robot navigation,image retrieval,3D reconstruction and have become the research directions of many experts and scholars.As image matching is affected by many unfavorable factors,it is necessary to study more robust features and more efficient matching methods in order to obtain better matching results.In the development of feature extraction technology,the proposed SIFT algorithm is of great significance.The SIFT algorithm establishes the scale space by convolving a set of Gaussian functions with the image,which solves the problem that detecting locations are changed due to scale change of the image.By assigning a consistent orientation to each keypoint based on local image properties,the descriptor can be represented relative to this orientation and therefore achieve invariance to image rotation.Many algorithms were proposed on the basis of SIFT algorithms,such as PCA-SIFT and ASIFT.KAZE algorithm is an image feature extraction algorithm with better performance than SIFT algorithm and its related algorithm.KAZE builds the scale space through nonlinear diffusion filtering,which effectively avoids the image edge blur and loss of detail caused by linear Gaussian filtering.This paper proposes Li-KAZE algorithm based on KAZE algorithm.Li-KAZE does not need to establish a reference direction for descriptors to achieve rotation invariance.Instead,the circle is used to re-divide the neighborhood to ensure that the information contained in each sub-area is still the same when the image rotates.At the same time,the second-order gradient information containing more image details is added to the descriptor to improve the distinctiveness of the feature.Li-KAZE uses a linear combination of distance and chessboard block distance to measure similarity between feature vectors.The simulation results show that the Li-KAZE algorithm is 27.3% less than the original algorithm in terms of operation time,and the Li-KAZE algorithm is close to the original algorithm in terms of correct matching rate.The paper proposes a new descriptor named the L-LDB descriptor for the AKAZE feature.Similar to the M-LDB descriptor,L-LDB descriptor also adds orientation and gradient information for LDB.The difference is that L-LDB uses the image centroid method to determine orientation,and in order to obtain distinctive features,the image's second-order gradient information is added to the description information.AKAZE-LLDB features can be generated by using the L-LDB to represent the features detected by the AKAZE algorithm.In the simulation experiments of L-LDB,M-LDB and I-LDB,the detection features are based on the AKAZE algorithm,and the feature description uses the three descriptors respectively.Experiments show that the L-LDB descriptor is superior to the M-LDB descriptor proposed in AKAZE algorithm and the I-LDB descriptor proposed by Wu Hanqian in the correct matching rate and matching time.Finally,this paper combines AKAZE-LLDB feature with GMS algorithm to form a complete feature extraction and matching method.This method uses the AKAZE-LLDB algorithm to detect features and generate binary form feature vectors.After extracting the features,the initial matching is performed by means of Brute Force algorithm,and the similarity measure adopts the Hamming distance.After the initial matching,the GMS matching is performed.GMS algorithm introduces motion smoothness constraint and eliminates the mismatch by transforming the constraint into statistical measure.The experimental results show that the AKAZE-LLDB +BF+GMS matching method reduces the matching time by 45.2% compared with the traditional AKAZE-LLDB +Ratio + RANSAC matching method and has better matching effect.In addition,the use of AKAZE-LLDB+BF+GMS matching method can effectively avoid the loss of the existing target tracking process,further verifies the practicability of the matching method.