Research On Image Corner Detection Algorithm

The corners of an image are very critical information when describing the characteristics of an object.The corner detection of an image is a pre-processing step for complex applications such as target detection,target tracking and image classification.The quality of corner detection directly affects the quality of subsequent image processing.Therefore,corner detection of images plays an irreplaceable role in the field of computer vision and image processing.It is of great practical significance to research and design a corner detection algorithm with high accuracy.This thesis focuses on the intensity-based corner detection algorithms and contour-based corner detection methods.In response to the shortcomings of traditional methods,two novel corner detection algorithms are proposed in this thesis.The performance of the algorithms is evaluated by different experiments which finally verified the effectiveness of the algorithm.The thesis can be summarized into two parts: 1.Corner detection algorithm based on bilateral filtering.This thesis proposes a corner detection algorithm that combines the first and second order derivatives of the image to construct a local structure tensor,then uses a bilateral filter to extract the gray-scale variation information of an image and to calculate the corner measurement.It improves the current problems,where multiple pseudo-corner points arise in the current intensity-based corner detection.Above all,the first and second order derivative information of the image are combined to calculate the local structure tensor,and construct the corner measure,which help to accurately distinguish edge points from corner points(to reduce the interference of the edge points),and to ensure the accuracy of the algorithm.Then,to address the deficiency of gaussian filter,bilateral filter is used to filter the local structure tensor.At the same time as filtering the noise,the feature information of image edge and corner is well preserved to ensure that the corner detection has a high response to the real corner,and effectively removes the pseudo corner.Finally,a new corner detection algorithm based on bilateral filtering is constructed using the proposed corner measure method.In this thesis,some common corner detection images are used to perform noise robustness experiments and corner matching experiments based on the algorithm.Four affine transformations were carried out to obtain 1050 test images based on 24 real world gray images of the different scenes.The average repeatability of the proposed algorithm under different image transformations was calculated and compared with other four classical corner detection algorithms.The experimental results show that the proposed algorithm has good detection performance.2.Contour-based corner detection based on the angle difference of principal directions.The traditional contour-based corner detection uses the local curvature information of the edge to construct corner measure for corner determination.This kind of algorithm is very sensitive to the noise and detail changes on the edge,which easily leads to the instability of detection results.In response to this problem,a contour-based corner detection algorithm via angle difference of principal directions is proposed in this thesis,combining with the gray information on the edge contour.First,the Canny edge detector is used to acquire the edge mapping in the image,then each edge line is extracted and the gap is filled.Then,the intensity changing information of vertical direction and horizontal direction at each point on the edge is calculated,and the principal direction of gradient change is found.Finally,corner detection is performed by constructing the corner measure using the angle differences of gradient principal direction between adjacent pixel points on the edge.In this thesis,classical test images of corner detection are used for corner matching experiments,and the test images selected from the standard data set are used to perform the average repetition rate experiment under different affine transformations.Experimental results show that the proposed algorithm performs well under different transformations.