Research On The Corner Detection Algorithms Based On Gray Change

Image feature extraction is a research hotspot and key technology in the field of computer vision and pattern recognition.As one of many features of the image,corner not only has the invariance of illumination,rotation,etc.,but also has rich information content and small amount of data,so it is widely used in tasks such as image matching,camera calibration,motion estimation,3D reconstruction and target recognition.In these tasks,corner detection is both a basic work and a key step,and the result of corner detection will directly affect the performance of subsequent image processing.Therefore,it is of great significance to study and improve the relevant theories and methods of corner detection.In this thesis,the corner detection algorithms of digital images are firstly studied.On this basis,aiming at the problems that the corner detection algorithms based on gray change are sensitive to scale changes,run slowly and have poor self-adaptability performance,the in-depth research is carried out.The specific research work is as follows:(1)The definition of corner,the evaluation criteria for the performance of the corner detection algorithms,the general process of corner detection and the related techniques used in the process,as well as the principle and its corner detection steps of several typical corner detection algorithms are introduced.(2)Aiming at the problem that the Harris algorithm is sensitive to scale changes and runs slowly,an improved Harris corner detection algorithm based on similar pixels is given.Inspired by SUSAN algorithm,the improved algorithm first calculated the number of pixels within the 8-neighborhood template of the target pixel that are similar to the target pixel,and selected candidate corners by its value.Then,the corner response function is improved by the number of similar pixels of candidate corners.Finally,the local non-maximal suppression is performed to determine the final corners.Experiments show that compared with the Harris algorithm,the improved algorithm shortens the corner detection time,and the corners extracted by the improved algorithm are more accurate and stable.The proposed algorithm can improve the corner detection efficiency and stability of Harris algorithm.(3)In order to remove false corners,reduce corner omission and detect the corners of the image in a real-time manner,a fast Harris corner detection algorithm based on gray difference and template is proposed.The improved algorithm selects initial corner set by the combination of gray scale differences of each pixel in the image and small template;On this basis,the SUSAN algorithm is improved and the improved SUSAN algorithm is used to refine the initial corner set.Finally,the corner response function values of the initial corners are calculated and the non-maximum suppression is performed to determine the finial corner.Experiments show that the corner detection time of the improved algorithm is significantly reduced and is only 4.7% that of the original Harris algorithm,which can satisfy the requirement of the real-time corner extraction in the images.(4)Aiming at the problem of poor self-adaptability and real-time performance of SUSAN corner detection algorithm,an adaptive and fast SUSAN corner detection algorithm is presented.Firstly,the improved algorithm calculates the gray difference of each pixel in the image at the four compass directions and selects candidate corners according to a screening threshold.In order to adaptively determine the screening threshold,the gray standard deviation of the image is selected as the screening threshold.Then,a circular template is used to traverse the candidate corners and the gray standard deviation of the pixels in the template is used as the difference threshold to realize the adaptive extraction of corners of SUSAN algorithm.Experiments show that the improved algorithm can extract self-adaptively corners in the image and reduce significantly the corner detection time.The improved algorithm enhances the self-adaptability and real-time performance of the SUSAN algorithm.