| 英文摘要 | The main contributions are two-fold:1. A practical coarse image curve matching method based on accumulation of evidence is proposed. The method can efficiently and robustly find the coarse location of a usually short extracted image curve into a long reference curve. The main characteristics of the proposed method are: Firstly, it is shown by experiments that the distance between two curve points is more reliable than the curve length itself to be used as a matching invariant in the presence of noise, hence in our work, the distance from the start point to the end point of the extracted image curve is primarily used for the selection of matching candidates in the reference curve. Secondly, the evidence accumulation concept is introduced in our matching algorithm, which not only significantly eliminates the incorrect matching of control points, but also substantially decreases the proportion of curve segments to be verified finally, a time consuming process. As a result, the computational efficiency and robustness of the proposed method is largely increased; Thirdly, the statistic Gaussian model is introduced in the classical Hausdorff distance calculation, which makes it more fit for the matching of curves with significant partial deformations. Extensive experiments with simulated data and images of satellite and digital camera show that our proposed curve matching method is practical, efficient and robust.2. A new planar curve matching method based on Fourier-Mellin Transform is proposed. In this method, the two curves to be matched are converted into two binary images at first, then the Fourier-Mellin based image registration method is used to register these two binary images to estimate the matching parameters of the two curves. The advantage of the proposed curve matching method is that it does not need extracting features nor establishing feature correspondences, the two difficult steps generally used in other curve matching methods. Instead it matches the two curves in frequency domain using their global similarity, and is inherently insensitive to random noise and local distortion. Experiments show that if the two curves to be matched are related by a similarity transformation, or its reasonable approximation, the proposed method can achieve satisfactory matching results. Besides, a concrete application, i.e., the coarse localization of remote sensing images, is reported to illustrate its applicability and usefulness.Besides the above curve matching problem, the present author also spent nearly two years on the selective attention problem, and proposed a “perceptual object” based selection model. Under this model, perceptual objects, which are defined as a connected region with homogenous gray level distribution, are hypothesized as the primary attention unit and an associate operator is proposed for computing the saliency of objects. |
修改评论