Research And Improvement On SIFT Feature Matching Algorithm

Image feature matching has become a hot research field in computer vision anddigital image processing, it’s primary task is to extract the feature points which haveimage important characteristics from the image and describe it, then through theestablishment of correspondence of feature point set between the original image and thematched image, calculate the similarity measure between the corresponding points todetermine whether the image is matching or not, or through the methods which need notto establish the correspondence to match feature points between images, such as theminimum mean square error matching, a quick search matching, Hausdorff distance andso on. The common matching algorithms include the spatial relationship based matchingalgorithm, the invariant descriptors based matching algorithm, pyramid and waveletalgorithm. David G.Lowe had proposed a point feature matching algorithm--SIFTalgorithm has invariance, quantities, unique features for various kinds oftransformations, such as scale, light, rotation, which has been applied to facerecognition, image retrieval, object matching and other fields.In most practical applications, due to unfavorable factors such as noise andvarious kind image transformation will bring the robustness affection for SIFT operatorwhen conducting features matching, it is directly related to the performance of thealgorithm. So, quick extract stable features to form an effective feature descriptor is amajor research goal in this article. We propose two improved methods for featuredescriptor from a different perspective：one is global structure of SIFT descriptor.through changing the neighborhood area of feature points, the robustness of thesub-global characterization and the anti-rotation scale invariant ability has beenstrenghened and the affection which various kinds of image transformation factor bringhas been reduced. The other one is rotation invariant textured SIFT descriptor, calculatethe rotation invariant texture features around every key feature points image area, thenfusion RIT(Rotation Invariant Textured)characterization into the SIFT descriptor vector,and more comprehensive description of the image information is formed, and thecomputational complexity of real-time image match with the reference image has beenreduced by using the completely rotation invariant of texture features, the establishionof the local texture feature descriptors provide a better fault-tolerant for featurematching with positioning errors, it reduces false match rate to a certain extent and improves the matching results.Firstly, this article describe all kinds of image features and three kinds of featureextraction operator, then introduce the basic concepts of the SIFT algorithm and the keytechnologies involved, the expansion algorithms of SIFT descriptor and the expansionalgorithms of matching are given, point out the advantages and disadvantages of variousalgorithms; Then due to the shortage of the current SIFT algorithm, we propose twoimproved SIFT algorithm: GS-SIFT and RIT-SIFT, narrate their descriptor generationand descriptor matching strategy respectively, finally, analysis the matchingperformance between the traditional algorithmin in the case of rotation, scale, lightchanges, perspective, the experimental results are given.