Research On Image Feature Points Detection And Matching

The image contains abundant physical information,including line feature,texture feature,edge feature and point feature,among which point feature is the most basic feature.Image feature point detection and matching technology has a very wide and important application in the field of image processing,which can be applied to image mosaic,object recognition,3D reconstruction and other fields.Since the development of image feature point detection technology,many classical algorithms have been proposed,such as SIFT,ASIFT and so on.Most of the advanced algorithms are improved on the existing classical algorithms,trying to improve the operation efficiency,realize the real-time operation or further increase the algorithm performance,and get more image feature points on the existing basis.In real life,most of the images have different angle tilt.Most of the existing image feature point detection algorithms can only be applied to small angle image tilt.When the tilt angle is slightly increased,the quality of the detected feature points is extremely poor.ASIFT algorithm is proposed to solve the problem of feature point detection in large angle oblique image.ASIFT algorithm effectively increases the number of feature points,but brings the cost of time.In this paper,the image feature point detection and matching technology are studied,and a new algorithm for feature point extraction of large angle oblique image is proposed.Based on SIFT algorithm,an efficient feature point detection method for oblique image is proposed directly.At the same time,the ASIFT algorithm is tried to further improve the quality of feature point detection and increase the number of detection.This paper attempts to combine the new algorithm with advanced image mosaic algorithm to get high quality image mosaic effect.The research work of this paper is divided into three parts:1.SIFT is a classic and still widely used algorithm in image feature point extraction algorithm.SIFT feature point extraction has strong robustness and invariance to illumination,rotation and scale.When the image tilt angle is too large,SIFT algorithm is not suitable because it can not extract enough feature points.In the third chapter of this paper,a new improved SIFT algorithm is proposed.Firstly,the large angle image to be matched is pre-processed to get the mapping relationship of feature points in the image to be matched.Next,through the obtained mapping relationship,the large angle image is transformed by positive angle matrix,and the feature points of the transformed image are extracted again.Under the condition of basically unchanged running time,the number of feature points extracted by the improved SIFT algorithm is greatly increased,which is several times higher than that of the SIFT algorithm.2.The classic ASIFT proposes an effective method to extract feature points in the case of large angle of view tilt,and the performance is greatly improved compared with SIFT.ASIFT mainly simulates the transformation of the relative longitude and latitude angle between the camera and the image in actual shooting,in which the affine transformation is actually a two-dimensional space transformation.In the fourth chapter,a new PSIFT algorithm is proposed based on ASIFT,which simulates the real three-dimensional space rotation through the camera imaging model and space rotation formula,and realizes the longitude and latitude angle transformation of camera shooting.Firstly,according to the principle of keyhole imaging,the image is placed in three-dimensional space,and then the image is rotated in three-dimensional space to get the space front view.Finally,the space front view is mapped to two-dimensional plane for feature point matching.Experiments show that compared with ASIFT algorithm,PSIFT can effectively increase the number of feature points,even double the number of feature points in some images,and can detect a large number of feature points in the region where ASIFT does not have feature points.3.The performance of PSIFT algorithm proposed in the fourth chapter is better than ASIFT algorithm,but the spatial rotation angle selection follows the idea of ASIFT algorithm,which will produce large angle images rarely seen in practical applications.Therefore,the fifth chapter of this paper attempts to improve the PSIFT algorithm to further reduce the time loss,increase the selection of horizontal angle of spatial rotation and reduce the selection of vertical angle.Compared with the PSIFT algorithm,the number of feature points of this algorithm is relatively reduced,but it effectively improves the calculation speed and saves more than half of the time.By combining the improved PSIFT method with the current advanced parallax-tolerant image stitching based on robust elastic warping,the accuracy of image mosaic is improved,and the fused image keeps clearer local details than the original algorithm.