Image feature detection and matching is a key problem in many applications of computer vision, such as object detection, 3D reconstruction, image registration and video understanding. Though great progress has been made in this filed recently, it is still a challenging problem, due to complexed imaging conditions, a large family of image scene types, various shape distortions. The main contributions of this dissertation can be summarized as follows: (1) By introducing the tractrix and pseudosphere in mathematics into the field of image processing, a novel image filter called the pseudosphere filter is presented. Besides a scale parameter, an edge-preserving parameter is introduced in the pseudosphere filter, and thus a better trade-off between image smoothing and edge locating can be obtained using it. A pseudosphere-based edge detector is formed by replacing the Gaussian filter in the classic Canny edge detector with the Pseudosphere filter. Compared with the classic Canny edge detector, in the case of having the same smoothness, the pseudosphere-based edge detector offers a better precision for edge locating. (2) The concept of gradient inner product energy is introduced, and it is proved mathematically the gradient inner product energy can overperform the gradient magnitude in restraining the noise and tiny edges. A novel image edge detector called the inner product energy-based edge detector is presented by replacing the gradient magnitude in the Canny edge detector with the inner product energy. Compared with the classic Canny edge detector, in the case of offering the equivalent precision for edge locating, the inner product energy-based edge detector performs better in de-noising and tiny edges controlling. (3) Focusing on the problem of corner localization, a novel algorithm for corner detection and localization is proposed, which is based on local orientation distribution (LOD). The LOD-based algorithm can provide higher localization accuracy and perform more robust than most popular detectors. (4) A novel ideal to automatically match lines based on line descriptors is presented. The main steps of constructing a line descriptor are following as: firstly, the line neighborhood is decomposed into several parallel line segments or overlapped sub-regions; then, a line description matrix is formed by selecting an image feature; finally, a line descriptor is obtained by computing the mean and standard deviation of the column vectors of the description matrix. Based on the two partition methods of the line neighbourhood and different features, four line descriptors are proposed for line matching in this paper. Besides, the MSLD line descriptor can be straightforwardly extended for a curve descriptor MSCD. Experiments show that these descriptors can be competent for line and curve automatic matching.
修改评论