Optical Remote Sensing Image Local Feature Extraction Technology And Its Application Research

The local feature extraction technology of image is the basis to solve many problems of computer vision, and the characteristic parameters extracted can reflect the essential nature of the image, which is the key to correctly interpret the image. The main reason for its widespread attention and research by scholars both at home and abroad is that the image is considered to be composed of different local areas in image processing, avoiding image segmentation. The process of local feature extraction is to firstly detect useful information in the image, and then use appropriate feature description method to construct feature vectors in order to characterize detected information.As the technology of remote sensing develops rapidly, local features have been widely used in the field of remote sensing image processing. This thesis focuses on the key technology, such as the theory of scale space, feature detection method and feature description method with the purpose of obtaining local feature, which is more consistent with human visual characteristic and stronger robustness.The local feature extraction techniques are mainly as follows:(1) The local feature extraction methods were systematically expounded, including the theory of scale space, the means of local feature detection and feature description. Among them, the theory of scale space is the basis of extracting local invariant feature. How to build scale space and determine the characteristic scale of the image was studied. The main ideas and implementation process of some mainstream feature detection and feature description was discussed. The performance of these methods was summarized and evaluated.(2) A local feature extraction method based on MS-Gabor filter was put forward according to FAST feature without scale invariance and rotation invariance on the basis of in-depth research on the local feature extraction methods. First, MS-Gabor filter was designed combining with the theory of scale space and the multi-channel and multi-resolution characteristics of Gabor function, scale space pyramid was built so that FAST feature can be detected by searching multi-characteristic scale of the image in scale space. And then the binary descriptor was constructed using a new sampling model, after that Hamming distance was used for feature matching and RANSAC algorithm was used to eliminate wrong matches for feature purification. Finally, the simulated image sequence under different scenarios was acquired by using a reference image for different degrees of rotation changes, scale changes, blur changes and brightness changes. The experiments were conducted with image sequence obtained and the evaluation indicators of repeatability and matching score by comparing the performance of different detectors and descriptors, the experiment results showed that compared with other algorithms, the method proposed had a better performance in terms of rotation changes, scale changes, blur changes and brightness changes.(3) The application of local feature was studied in optical remote sensing image registration. The algorithm given in the thesis was respectively applied in image registration from the same sensor and different sensor. Further, the proposed algorithm was verified to be with considerable practicability by calculating registration accuracy.