| Image matching plays an important role in computer vision field, and is taken as a key technology in applying vision in other fields, such as in medical image detection, microscopic image analysis, content-based image retrieval etc. Besides, Image matching is also widely applied in the following fields, like fingerprint identification, face recognition, iris recognition, intelligent transportation management, motor vehicle detection, license plate identification in places like parking management centers, and it also can be used to analyze the behavior of the passers-by automatically in important places such as safety supervision department and be equipped in robots to identify tools, components intelligently in automatic production line.In recent years, the matching method, based on local characteristics, has become mainstream of the research on image matching, which extracts the interesting parts in the image as characteristics, generally requiring the characteristics in the image to remain stable after some kind of transformation.The common methods, on the basis of local characteristics, are as follows: shape context, steerable filters,PCA-SIFT,differential invariants, SIFT, moment invariants. Among them, SIFT (Scale Invariant Feature Transform) is a method put forward by Lowe by using the theory of Scale-Space.Krystian carried out an experiment to compare ten construction methods with invariant characteristics and evaluated their performance. The results showed that SIFT had the best comprehensive performance. In the article by Lowe, he pointed out two shortcomings in the SIF:1) When matched, the geometry layout among the characteristic points had not been used comprehensively;2) Matching results were susceptible to the effect of matching parameters. When matching conditions was loosely constrainted, it was easy to lead to the wrong matching, and when matching conditions were strictly controlled, it would tend to miss the correct characteristic point pairs of the matching.With regard to the first problem, Xia Shengping, on the basis of the basic description of SIFT vector matching, adopted procrustes iteration method to delete the wrong matching characteristic point pairs in the original SIFT. Liu Jianjun, based on Xia ShengPing’s study, used the procrustes iteration method twice to further improve the accuracy of the match.However, there still existed mistaken and missing matchings. The reason was that the procrustes methods only used feature point coordinates to do correlation calculation, the spatial relationship and information among the characteristic point groups had not been completely utilizedTherefore, the paper, taking the matching problems in SIFT characteristic points and the spatial construct and information into consideration, makes an attempt to combine strong gravitational field method and convex hull method to improve the matching effect among SIFT feature points. Theoretical analysis and experiments will be carried out to verify the feasibility and effectiveness the proposed algorithm. This paper consists of the following parts:(1) Image feature points extraction and elementary matchingImage feature point extraction is an essential step in image point pattern matching, which is the basis of the effective image matching. It will have a direct effect on the degree of the accuracy of image matching.Sift (scale invariant feature transform) local characteristics description operator is a method put forward by Professor Lowe to extract feature points. The feature points extracted by SIFT method possess the following advantages:the scale of rotating, zoom, brightness changes remain constant, even for the perspective change, affine transformation, and noise, they still have good stability, which makes it widely used in image identification field. In the paper, the author will use SIFT methods to extract feature points, and use Professor Lowe’s elementary matching methods to carry out the elementary matching.(2) Combining gravitational field method and convex hull method.In order to get more accurate image feature matching results, the author adopts gravitational field method and convex hull theory to combine the matching problems in SIFT characteristic points and the spatial construct and information.Convex hull is a concept introduced from physics representing the relationships between the points in the space. The gravity among the points embraces abundant information which can be used to solve the matching problems among the points. In the paper, the author only uses the scalar concept of gravitational field to make a quantitative description of the relationship among the points. Convex hull is a stable space structure, the points on the convex hull, even under the condition of changing image, can still possess a good stability, which, therefore, can be used in point pattern matching. In the paper, the author combines the two theories to solve point mode matching problem under the circumstance of changing RTS in an effective way, and even under the condition of partial affine transformation, certain effects can still be achieved(3) SVD (singular value decomposition) methodIn view of great changes in the affine changes, good results cannot be achieved. The authors attempts at using the singular value decomposition method to solve the affine problem. The singular value decomposition method can effectively eliminate the effects the crosswise changes and scale transform and translation have on point set, which only keeps component in rotation and greatly simplify the matching problem. Therefore, so far, it is a newly-made point pattern matching method. In this paper, the author first decompose SVD in the target image feature points matrix, which will lead to problems in RTS transform category, and then combines the above mentioned matching methods, the matching problem of affine transformation will be effectively solved.(4) The experiment and analysis resultsIn this paper, parts of images of Columbia University "Coil-100" data will be used to carry out the simulation experiment. Besides, the author selects several items to shoot, and obtain the target images based on similar transformation and affine projection transformation respectively to carry out identification test. In a similar transformation, each template corresponds with72groups of different transforming images, while in the affine projection transformation, each template on the basis of the similar transformation, combines with two groups of beveling parameters to produce images. This paper makes an attempt to compare this method with other SIFT point pattern matching algorithms and makes an analysis the research results.The theoretical analysis and experiment results show that based on the convex hull theory and strong gravitational field method, even under the circumstances of changing RTS, the matching problems among the SIFT characteristic points has been solved effectively. In the future, great efforts still need to be made to strive for better matching effect. |
PDF Full Text Request