A corner detection method based on strong noise adaptation
Yang Dong,Yin Yi-Long,Zhu Ming-Ying,Cao Qing-Feng
DOI: https://doi.org/10.3321/j.issn:0469-5097.2008.02.005
2008-01-01
Abstract:Corner is the point which has the sharp change of brightness in image,or the maximum curvature on contour.Accordingly,there are two kinds of corner detectors,contour-based ones and image-based ones.Contour-based corner detection plays an important part in shape recognition.Generally,there are three kinds of contour-based corner detectors,one based on maximum curvature,one based on beeline or curve fitting and the other one based on multiple scale spaces.However, up to now most of them don't work well under strong noise conditions.Under strong noise conditions,we suppose the average and uniformity of noise are unknown.It is mainly because of complex vision conditions,such as camera noise,complex illumination and mechanical vibration of camera.To detect corner adaptively when noise conditions are unknown,we can simulate human vision system.According to Gestalt vision theory,human vision is of good continuity,and can filtrate continuous noise details adaptively and gain glancing corners of contour.When human vision searches corners,it can neither compute the curvature of every point,nor find the optimal criterion globally.Someone points out that human vision is sensitive to beeline section.And the least square method(LSM) is widely used for beeline fitting.However,most of beeline-fitting-based corner detectors cannot work well when noise level changes.The main reason is that they assume the noise level is small or fixed and known,so they use a user-defined or fixed threshold to detect corners.This paper points out that a corner is the point or points-set,the forgoing and following points-set of which are most fitted with two beeline sections.And we use the whole points of local contour to detect a corner.Therefore,if the detail noise increases,it can only reduce its contribution to corner eigenvalue.There is the shortest distance between corners(SDBC),under which the two corners can be recognized because of the strong noise.In the head and tail of a contour there are points of SDBC,which cannot be detected as corners.The other points are called candidate points.We compute the residual sum of squares(RSS) of forgoing and following points-set of every candidate point as the eigenvalue.The eigenvalues of all candidate points are shown as an eigenvector curvature.If the local contour contains a corner,the eigenvector curvature will be an smooth concave,and the corner is located on the position of minimum value.If not,the eigenvector curvature will be a irregular plane line.Meanwhile,it uses the half-division searching to search the minimum value,and check whether the curvature is concave.This step can reduce the computation complexity.Starting from some points of a contour,partial contour with increasing length will be checked.The initial length and increased length will be SDBC.If a corner is found,the length increasing will be stopped and this corner will be the next start point.Both the theoretical analysis and the experiment results indicate the validity of the method in increasing the veracity and reducing the time complexity of corner recognition of monotone edge on strong noise level.