Research Of Image Registration Based On Video Features

Image registration is an important research area of computer vision and patternrecognition. It’s main objective is to determine the spatial position relation betweendifferent images of the same object or scene，which are obtained from different time,different perspectives, different sensors or different shooting conditions. Image registrationtechnique is widely applied in the field of aerial video sequence processing, medical imageanalysis, movement analysis and so on. Now, Image registration methods can be dividedinto two categories, manual registration method and automatic registration method.Automatic registration method is the ultimate goal of image registration technique.Automatic registration methods also can be divided into two major categories, grayinformation-based and feture based. At present, the registration method based on featurebecome the main research direction of image registration technique, because it has manyadvantages, such as small amount of calculation, strong robustness to the noise interferenceand the complex geometric transformation, etc.The change of gray, geometric, and the complexity of the scene often exist betweenthe different frames of the aerial video sequences; in addition to the overlap region betweensome frames are small, which makes it difficult to get the desired results in the registration.If only to improve registration algorithm in one part, it is difficult to achieve accurate result.To solve these problems, this paper proposes a new feature point-based image registrationalgorithm. By improving the detection method and the matching method of the featurepoints, the problem of register different frames of an aerial video sequence was solved.In the aspect of feature point detection, Harris-Laplace operator can be applied to graychange, complex geometric deformation, and noise, etc. It’s a good detection operator.However, the feature points detected by the operator often contain redundant points andhave the problem of distribution concentrated, which will affect the accuracy of registrationand have difficulty in the registration of pictures with small overlapping areas.Aiming atthese problems, this paper improves Harris-Laplace operator by combining the informationentropy and sub-block strategy. The improved operator can not only extract more evenlydistributed feature points and avoid excessive detected feature points in the flat area, butalso remove redundant points effectively.In the aspect of feature point matching, by combining the extended Hu moments with the scale information of feature points, using the strategy of bidirectional matching, theaccuracy of the initial matching points was enhanced and improve the efficiency of thealgorithm and the accuracy of registration were improved.Experiments show that, through the improvement of the above two aspects, the newregistration method based on feature points in this paper can achieve a good effect in theregistration of the different frames of the aerial video sequences. In the end of this article,aiming at the problem that through different DVD play video will exist migration ofdifferent degree which was found in the experiment, using a simple registration method,accurate video offset was obtained.