Study On Image Invariant Feature Detection And Description

Feature detection and description are fundamental tasks of many image procession and computer vision. The performance of feature detectors and descriptors directly determines the efficiency and precision of image procession. It is important that image local feature should be as distinctive as possible while also be robust to occlusion, background clutter and noise, invariant to various image transformations due to translation,rotation,scale,affine deformation, difference in illumination, object movement, and change in viewpoint. Recently the study of image invariance feature based upon the invariant theory of computer vision has been one of the most active researches in the filed of image procession. To these difficulties, this paper is made several matters following:Harris corner detection is a classical algorithm, but hasn't the property of scale invariant. In this paper, the multi-resolution idea is introduced into the Harris algorithm, and the wavelet-based formula denoting the image intensity variation is constructed, meanwhile, the auto-correlation matrix is obtained that reflected the scale variation information. Then, a novel multi-scale Harris corner detection algorithm based on wavelet is presented, which can detect the corners in different scales and overcome the drawback that the single-scale Harris detector usually leads to either missing significant corners or detecting false corners due to noise. Compared with Harris detector, the presented algorithm is more efficient in corner locating and is robust with respect to noise.Local feature descriptors should be distinctive, robust to occlusion, noise and so on. Recent work has concentrated on making these descriptors invariant to image transformations. In this paper, a novel curvature feature vector descriptor (CFVD) is presented. For each feature, the curvature feature vector accumulates curvature values rather than counts edge points in each 5×12 log-polar coordinate system. Each pixel's curvature value is weighted by a Gaussian window and then added to the corresponding bin of a 3D histogram. Therefore the feature descriptor is a 5×12 vector. This new technique combines the Shape Context with SIFT and produces a feature descriptor, which is based on global information, but doesn't depend on edge or contour extraction. So it is robust to local appearance similarity, noise and non-rigid transformations. The match experimental results via theχ2similarity metric illustrate the algorithm has a good performance of accuracy and efficiency.To overcome the drawback that the existing registration algorithms only suitable to the small rotation angle between images (about 0~5°), in this paper, a novel gradient orientation histogram is built using a Gaussian-weighted circular window and a major orientation to each feature point is assigned based on local image properties. Then the orientation difference between two images is calculated from an angle histogram, which is presented with feature point major orientations instead of former gradient orientations. Finally, a novel registration algorithm, which is invariance to rotation and robust to noise, is obtained. The experiments show that this approach has no restriction to the rotation angle between the images and the registration results are satisfying.At last, experiments and analysis are given to show the algorithms we proposed have a good invariance and application performence in improved registration algorithm.