Based On Scale Space Of Local Invariant Feature Extraction And Matching Algorithms

Image feature extraction is an important research content in computer visionresearch, and also a research of a lot of problems. Because of the target imagebetween most of the point, scale, rotation, light, and the fuzzy transform, so thestability of how to extract the image feature is related to areas of research focus. Inrecent years, a kind of local invariant features due to its for image translation, choosecanary，scale, lighting and perspective transformation invariance，has been widelyused in many fields. Method based on local invariant features of the main steps offeature extraction and feature matching. This paper deeply analyzes the invariantfeatures of related theory knowledge, studies the existing local invariant featureextraction algorithm and the matching strategy, aiming at the problems of the currentis improved, the improved method can get better effect.In2004by Lowe SIFT image matching algorithm is put forward, has strongrobustness and stability, and has the very high distinguish rate, and128-dimensionalspace feature descriptor has very good in dependence. However, it also contains a lotof redundant information. Based on SIFT algorithm proposed a classic based on sobeledge detection with K-L transform efficient SIFT algorithm; First using sobel edgedetection operator, by setting the threshold T，ignoring some redundancy featurepoints, reduce redundancy feature vector is generated; Secondly, by K-Ltransformation, cut down the128-dimensional space feature descriptor to60d，reduce the complexity of time; Experiments show that the higher the threshold value is set, the less access to the key point, the higher the efficiency of the match, visiblealgorithm proposed in this paper as a whole to achieve the real time and highefficiency of matching. For image matching in the application of the embeddedterminal situation of high real-time requirements, this article also puts forward animproved SIFT algorithm, key points in traditional SIFT algorithm is stable, the keyto make use of contour let transform frequency domain global texture description, andthe global texture similarity calculation results to sort, select the top1%of points, andthen through a kind of based on Vector Angle (Vector Angle, VA) the approximatenearest neighbor search algorithm matching, improve the matching accuracy andspeed, improve the accuracy of the embedded terminal.Scale Harris method detected feature points there are a lot of redundant points,although Harris-Laplace method can remove some redundant, but there will be a localstructure within the existence of multiple feature points or a feature point on behalf ofthe local structure of multiple scales. Therefore, an improved method in detectingmulti-scale Harris features point tracking group. That represent the same localstructure is divided into a set of feature points, using normalization of Laplacefunction to remove redundancy, recycling point of measurement to select the bestrepresents the characteristics of the structure of the local point. The experimentalresults show that the method can effectively remove the redundant points, when thefuzzy and rotation transformation is better than Harris-Laplace method, has thescale-invariant features.Based on the affine Gaussian scale space theory, this paper proposes a fully affineinvariant feature extraction algorithm (FAIF).FAIF algorithm aiming at the problemof affine Gaussian scale space is difficult to construct, Gaussian scale space aiffnetransformation is put forward for the thinking of scale space. With the area of theimage characteristics of covariance matrix as the feature area anisotropy degree ofmeasurement, the anisotropic characteristics of the area by rotating compressiontransformation for isotropic region, finally on the isotropic area completely affine invariant feature point extraction. The experimental results show that the FAIFalgorithm can adapt to the large Angle and scale transformation, and also in stereoimage matching point enough，its performance is superior to the existing afifneinvariant feature extraction algorithm is proposed.