Local Feature Extraction And Its Application In Image Registration

When the background is not clear or cluttered,the global features of the image are more sensitive.At this time,the information extraction and recognition of images are seriously disturbed.Because of the uncertainty of foreground and background information,these problems often make the limitations of global information of image gradually appear.Image registration requires that the local feature descriptor of the extracted image has good feature representation ability and remarkable discrimination,which makes the matching accuracy of the same-name point pair higher and the number of the same-name point pairs more.Based on the above understanding,a new multi-directional multi-scale feature representation method for local feature description of images is proposed.The concept of multi-directional multi-scale feature cyclic vector and the similarity calculation method of multi-directional multi-scale features are defined,and an image registration framework using multi-directional multi-scale features is developed.The algorithm firstly extracts corners to reduce redundant information in the image;secondly,extracts multi-directional and multi-scale features at the corners of the image with abundant information as descriptors of the local region of the image;and then,eliminates a large number of non-representations between the source image and the target image by calculating similarity matrix and maximum value search.Matching points;finally,the RANSAC algorithm is used to reduce mismatched matching points.Gabor filter banks with multi-scale and multi-direction are selected in the experiment,and the experiment contents are divided into various satellite image registration experiments under normal conditions and various satellite image registration experiments with large changes in external environment.The 1-to-1 matching experiment of candidate homonym pairs shows the uniqueness of multi-directional and multi-scale features,while the 1-to-1 matching experiment of homonym pairs shows the rotation invariance of multi-directional and multi-scale features.In all the registration experiments,there was no mismatch except for one mismatch of the same name point detected by SIFT algorithm.In the case of conventional image registration,the number of points with the same name is more than SIFT algorithm,and some even more than double.For severe registration conditions caused by earthquakes,tsunamis and wars,the number of points with the same name not only exceeds SIFT algorithm,but also differs by more than one order of magnitude at most.Experimental results of image registration show that the multi-directional and multi-scale features extracted from different types of images have good feature representation ability and significant discriminant power.The proposed algorithm has translation invariance,rotation invariance,scale invariance and a certain degree of illumination invariance.It is suitable for image registration in many scenes.Compared with SIFT algorithm,most of the cases have more logarithms of the same name and lower false matching rate.The disadvantage of the algorithm is the long running time caused by the huge amount of computation.The solution is to study the related parallel computation and fast algorithm and adopt other filters with less computation.In the future,it is believed that the application of multi-directional and multi-scale features extracted by other filters in image registration will be an important research topic.