Research And Realization Of Image Registration Based On Improved SIFT Algorithm

Image registration technique is widely used in many fields of society andeconomy, such as medical, remote sensing, industrial, military, agriculture, etc. Imageregistration can be roughly divided into two major categories: manual registration andautomatic registration. In recent years, the automatic image registration has becomethe hotspot in this field. Automatic image registration can be divided into two aspects:based on region and based on the characteristics. The former uses gray level andlocation information of regional pixel directly, compute for similarity of sensed imageand reference image, obtain transformation model parameters and realizes theregistration finally; The latter need extracting image features and features matching,then obtain transformation model parameters according to the result of featuresmatching, which can achieve more robust results of image registration. Imageregistration technology based on the characteristics of the overall performance isbetter than that based on the region, so it has got extensive attention to scholars, andhas obtained many research results.Feature extraction and feature matching are the main research contents of imageregistration techniques based on image feature, and also the central issues ofalgorithm improvement. As a result, the research of image feature representation hasthe very vital significance. Image features usually include shape features, colorfeatures and texture features. Different features have different methods of featurerepresentation. From another perspective, these features respectively divided intoglobal features and local features. SIFT algorithm is one of the best algorithm basedon corner feature extraction, which is the most robust. SIFT features belong to thelocal characteristics, and has a good scale, translation, illumination and affineinvariance. Due to its good performance, image registration technique based on theSIFT algorithm is getting more and more focused.At first, this paper discusses the basic theory of features representation, and thebasic theory and methods of image registration; And then compare image registrationmethods based on the angular point feature extraction algorithm (Harris, SUSAN and SIFT); finally chose the best robust SIFT algorithm, and put forward an improvedimage registration method based on SIFT algorithm: to reduce the time complexity offeature extraction combining with the Harris algorithm; to reduce the feature vectordimensionality combining with the Principal Component Analysis. Through theexperiment, it has achieved good results.