A Research On LATCH Image Matching Algorithm Combined With SURF

As one of the key technologies of computer vision,image matching has been widely used in various application fields of computer vision.Especially in object recognition,image registration and image retrieval,image matching is their core technology.The local invariant features of the image have the characteristics of invariance,redundancy and uniqueness under various image transformations(such as geometric transformation,illumination transformation,etc.),and are widely used in the field of image matching.In the problems of image matching based on local features,local feature extraction and feature description are very important steps.The detection algorithm and descriptor performance directly determine the efficiency and accuracy of image matching.Therefore,the image matching algorithm based on local features has great research value,and this thesis discusses and studies the image matching algorithm based on local features.In this thesis,the speed of image matching based on the binary descriptors is faster than floating point number descriptors,an image matching algorithm based on the Speeded up robust features(SURF)algorithm and the Learned Arrangements of Three Patch Codes(LATCH)algorithm is proposed.The main work includes: 1)For the problem that LATCH descriptor does not have scale invariance and rotation invariance,this thesis detects the main direction and scale information of feature points according to SURF algorithm,and repositions the triplet pixel block when LATCH calculates descriptors.Then binary descriptors with certain scale invariance and rotation invariance are presented.2)In this thesis,the problem of noise in feature localization is studied.The projection error of the feature point pair is calculated by the Random Sample Consensus(RANSAC)algorithm.A method of characterizing the error distribution using the circle centered on the feature point is proposed.3)According to the distribution of feature localization noise,this thesis proposes an algorithm that uses the shortest distance from the point to the region to calculate the projection error and then judge the inner and outer points.4)This thesis uses the Oxford dataset to evaluate the performance of the binary description algorithm and the performance of the optimized internal and external point determination algorithm.The experimental results show that on the Oxford dataset,the feature detection and feature extraction algorithm combined with SURF and LATCH is shortened by about 4.5% compared with the SURF algorithm.The binary descriptors in this thesis show better scale invariance and rotation invariance on the Boat image set than the original LATCH algorithm.Although the optimized internal and external point determination algorithm increases the running time by about 6.7% compared with the RANSAC algorithm on the Oxford dataset,however,in the case where the NSE values of the algorithm are almost equal,the number of interior points detected by the algorithm in the first image and the second image in the Graffiti image set is increased by 63.In the first image and the third image,the number of interior points detected by the algorithm is increased by 14.In summary,the scale invariance and rotation invariance of feature descriptors are added based on the LATCH algorithm.The optimized internal and external point decision algorithm also shows better performance.