Gradient Magnitude Weighting Of Descriptors And Main Orientation Design For Multispectral Images

Due to the development of multi-spectral image technology and the cost reduction of the image equipment,more and more multi-spectral imaging equipment has been widely used.These images from different sources have different characteristics and they play an important role in remote sensing,public security,traffic,fire control and other fields.The information of multi-source image contains complementary and redundant information.Therefore,it has received more and more attention.In this paper,firstly,we introduce the basic principle,method and step of image registration algorithm,and the aporia of multi-spectral image registration.Then we start from these difficulties,and explore the improvement scheme of the algorithm,which is based on gradient magnitude weighting of descriptors and design of main orientation.Finally,through the database of a large number of multi-spectral images,we evaluate and compare the image matching effect of the proposed algorithm.Multi-spectral image registration technology has the following difficulties:the intensity and gradient magnitude of the corresponding pixel has not linear relationship,the detection rate of the keypoints is low,and the gradient direction may appear to reverse.In this paper,we propose a gradient magnitude improvement algorithm based on edge pixels which is stable in the multi-spectral images,we use these edge pixels to compute the gradient histogram of each feature point.On the other hand,for the low repetition rate of the feature point detection and gradient direction inversion,we propose an improved feature matching algorithm,which consist of the traditional SIFT feature points and the PIIFD main direction,and we also can use the line segment intersections as the feature points,and the orientation which is located in the first or the second quadrant of the first line as the main direction of the feature points,in order to improve the robustness of the image rotation transform.Finally,we use a large amount of multi-spectral image data to make experiments,based on Studio Visual 2012 and Matlab platform,then we get the matching accuracy statistics,compare with the other traditional registration algorithms,and make summary conclusion.The experiments results show that the gradient magnitude weighting algorithm based on edge pixels which make use of the stable edge pixels of the multi-spectral images and retain the original gradient information could improve the registration effect of the original algorithm.Using line segment intersections as feature points can increase the number of feature points,and improve the detection rate of the feature points.The main direction of the intersection of line NO.1 also can enhance the robustness of the image rotation.In the case of the image with large angle rotation,it also has good registration effect.