Research On Key Techniques In Image Feature Point Matching

Machine vision is one of the important means for computer to percept the world.With the continuous development of image processing technology,the ability and scope of humans to acquire and process images are constantly increasing.With the advent of big data,artificial intelligence and the "Internet +" era,how to effectively process and use the massive amounts of images to better serve the development of human society is a challenge and opportunity in the field of image processing research.Image feature matching is a basis for various image processing techniques.The research focus of image feature point matching technology includes feature point detection,feature point description,feature point matching and removal of mismatches.The main work of this paper focuses on two aspects: feature point description and removal of mismatches.One key point of this thesis is to introduce the invariant moment theory into the description of image feature points to obtain affine invariant feature descriptors.The second focus of this thesis is to remove the image feature point mismatches by precise shape context features and belief propagation probability inference network to further improve the matching accuracy.The specific work is as follows:(1)The Zernike moment is introduced into the image affine invariant feature descripting.The Zernike moment itself has rotation invariance.The translation invariance of Zernike moment has been obtained by centroid translation.In order to obtain feature descriptors with affine invariant properties,the principal component analysis method is used to normalize the elliptical neighborhood obtained by the MSER algorithm into a perfect circle with equal areas.Then the Zernike moment of the feature point is calculated on the unit circle as the descriptor of the feature point.Experimental results show that compared with the previous Zernike moment descriptors,our method can obtain affine invariant image feature descriptors.At the same time,the fast pseudo-Zernike moment algorithm was ammended for the rest two quadrants,and the adviced length of the peoposed Zernike moment descriptor was determined through experiments.(2)The MSA moment was introduced for image feature descripting.The MSA moments theoretically are complete affine invariant.Introducing the MSA moment into the image feature descripting,in theory,the complete affine invariant feature descriptors can be obtained.However,the MSA moment calculation process was found to be abstract and complex,and the proof of the MSA moment was found to have a defect by thoretical reasoning and experiments.For these two problems,this thesis firstly gave concise geometric proof,and secondly corrected the defect of the original MSA moment.Then combined with the MSERalgorithm,the MSA moment was introduced into the image affine invariant feature description,and the true affine invariant feature descriptor was obtained.(3)The concept of extended contour is proposed,and the affine invariance of the extended contour is proved by theoretical reasoning and simulation experiments.The fast affine invariant feature theory has complete affine invariant properties,but if the integral operation is performed on the original image during the calculation,the affine invariant ability will be reduced because the integral domain was not affine invariant.After the extended contour proposed in this thesis was used as the domain of fast affine invariant features,the classification and recognition ability of fast affine invariant features has been greatly improved.At the same time,the concept of extended contour proposed in this paper can broaden the thinking of affine invariant feature extraction method.(4)The precise shape feature of the feature point and the belief propagation probability inference network were introduced into the removal of feature point mismatches.Firstly,this thesis uses the position information,scale information,direction information and adjoining information of the feature points to design the accurate shape features of the feature points,and the accurate shape features were used to remove the feature point mismatches.Secondly,the location information,scale information,direction information and adjacent information of feature points are integrated into the definition of evidence function and compatibility function of the belief propagation network,and the mismatches of feature points were removed by the belief propagation network.Simulation experiments show that the proposed two algorithms can remove the mismatches of feature points with high recall rate and accuracy.(5)According to the research content of this thesis,an affine image registration system was developed.The system includes all the affine invariant feature descripting algorithms proposed in this thesis.For comparison,the SIFT algorithm is also included in this system.The system's user interface is simple and friendly,and has strong scalability.The developed affine image registration system can be used as a reference or development basis for subsequent research.