Corner Detection Algorithm Based On Log-Gabor Filter And Its Application

Corner detection and description has always been the most popular research direction in the field of computer vision.It is the front-end processing part of the feature-based image understanding system,so its performance has a great impact on the subsequent processing of the system and the overall architecture.Nowadays,the information contained in twodimensional images is more and more abundant,but many existing corner detection algorithms are based on single-scale information,which will cause inaccurate detection of corner position information and even identify a large number of wrong corners.When it encounters interference such as image distortion,it will also affect the performance.Aiming at this situation,this thesis proposes a corner detection algorithm based on Log-Gabor filter.And based on the algorithm,a description is added for its corner points,which strengthens the practical application of the algorithm.Later,through different types of experiments,the effectiveness,repeatability and robustness of the algorithm were verified,which proved that the algorithm has better performance.Aiming at the problems of low positioning accuracy of traditional corner detection algorithms and a large number of false and missed corners,a corner detection algorithm based on Log-Gabor filter is proposed.First,we use the two-dimensional Log-Gabor filter bank to process the image information,and get the filtering results of the image in different scales and different directions.Then we do the candidate corner points selection.For each pixel,we derive and construct a multi-directional structure tensor at different scales based on the filtered results of the pixels contained in the point and its adjacent area.And we solve the corner response function value of this point on a certain scale.Then we compare it with its adjacent points.If the point is the maximum point in this area,then compare it with the defined threshold.If it is greater than the threshold,we mark the point as a candidate corner.Finally,we do the corner points selection.We judge the response values of the candidate corner points under other scales,if all of them are maximum values and are higher than the threshold,then the candidate corner point is considered to be a corner point.We conduct different experiments to verify the algorithm through different types of image libraries.Firstly,in the affine transformation experiment,we compare the algorithm in this article with11 classic algorithms.No matter in image rotation,scaling,quality compression,noise interference,we get a higher corner average than other algorithms.The repetition rate reflects the good repeatability of the algorithm in this paper.Secondly,in the corner matching verification and noise interference experiments,the number of false corner points and missed corner points of the algorithm in this thesis are less than those of other algorithms,which proves that the algorithm in this thesis has excellent robustness.Since the above algorithm only detects the position information of the corner points,it does no description of the pixel information.Therefore,we combine the descriptor of the SIFT algorithm and propose a descriptor algorithm based on the corner detection in this thesis We describe the detected corner information,and use this algorithm in subsequent corner matching experiments.According to the position and scale information of the corner point,the main direction of the gradient of the corner point is calculated first,the coordinate axis is established with the corner point as the center point,and the coordinate axis is rotated to the main direction.Then select the surrounding 4x4 sub-region modules,and then calculate the main direction of each sub-region separately,and use this information to finally form a128-dimensional descriptor.Through description matching experiment and comparison experiment,the validity and repeatability of this algorithm are verified.The detection and description of six groups of different transformed images in description matching experiment verify the effectiveness of the algorithm.In the comparison experiment,we compare our algorithm with the SURF(Speeded-Up Robust Features)algorithm.The number of successful matches in this article is higher than that of the SURF algorithm under various conditions,and as the degree of interference increases,the performance of this algorithm has always been better than the SURF algorithm,which verifies that the algorithm in this thesis has better repeatability.