Research On Image Registration And Application Based On The Local Feature Points

The image registration technique is one of important component of digitalimage processing, which for decades had been developing the essential andindispensable technology in the image understanding field and the deepprocessing of computer vision. The image registration falls into two generalcategories: the algorithms based on gray scale and the algorithms based onfeature information. The matching speed of the latter has advantages over theformer, especially the algorithms that are based on local feature information,which have high feature extraction accuracy and are robust to noise, distortion,illumination change, etc. Therefore, the algorithm based on local featureinformation has become a main focus within the internet academic circle.The algorithms based on local feature information are mainly studied, andthe algorithm based on ORB is the research priority. The main works are asfollows.（1）The basic theories of the algorithms based on feature informationare researched, including the local points’ extraction, the spatial transformationmodel, the local feature points’ description, the local feature points’ matching,purification of matching point-pairs, etc.（2）The research of ORB algorithm is focused on, including the ORBextraction algorithm of local feature points and the ORB description algorithmof local feature points. The experiments of the algorithm based on ORB areemployed to compare with the SIFT algorithm and SURF algorithm in theaspect of matching speed and matching precision, respectively. Meanwhile, anevaluation method of the image registration based on the difference imagemethod is presented, which is employed to compare the precision differenceamong them in quantity.（3）Under the background of defect detection of printing industry, thefast image registration algorithm based on ORB is employed to solve the problem, which is that the SIFT and SURF algorithm have slow speed and arenot suitable to the online detecting. The reason that ORB algorithm has nothigh registration accuracy is further studied, and the improvement is performed,which is that the projection transform model is replaced by the affinetransformation model to estimate the parameter matrix. The experiments areemployed to compare the registration accuracy and validate the validity of theimproved algorithm.