Scene Matching Based On Feature Points

Scene matching is an important image analysis and processing technique, and is used in a wide-range areas such as navigation and location, weather prediction, environmental studies, change detection and other fields.The difficulty of Scene matching technology is imaging conditions are subject to different sensors, different natural conditions, as well as the impact of various imaging distortion. How to design high adaptable, high precision and fast matching algorithm is the kernel study points. In this paper, scene matching is based on feature points.Based on analysis of the classification of existing scene matching, Concrete realization method of SIFT algorithm is first related. SIFT includes four parts: establishment of DOG scale space, using non maximum suppression to extreme point ,descriptor generation, and using kd-tree and best bin first to measure the distance.As a result of the existing of the false matching pairs, the relationship between adjacent feature points is used toEliminate false matches. Experiments shows that SIFT algorithm is invariant to scale transformation, illumination change, rotation, noise and other factors.Based on the analysis of the PCA theory in mathematical description, the specific studies of the PCA is applied to the generation of SIFT descriptor on the purpose of reducing descriptor's dimension. The experimental results show that, under the premise of no affecting matching performance, PCA-SIFT increase computing speed Substantially. The paper studies SURF algorithm, which has similar ideas with SIFT, but it is computed on the integral image, SURF uses several approximations, such as box filter in the process of establishment of scale space, fast Hessian in the process of detect key points, Haar wavelet when the descriptors are generated, and fast index in the process of similarity measures, The experiments show that SURF greatly increase computing speed.Finally, making experiment with SIFT, PCA-SIFT and SURF, and remote sensing image experimental data is compared and analyzed in detail.