Local Affine Invariant Feature For Automatic Optics Remote Sensing Image Registration

Compared with other features, affine invariant features are invariant to the viewpoint changing and camera parameter variance, and have great advantage over robust, repeatability, distinction and applicability, hence the theory and method of affine invariant feature extraction have became an active and challenging field. Taking these methods into remote sensing image automatic registration can significantly improve the performance of many fields, such as territorial resource exploration, ground mapping, military geology, etc. In this article feature extraction and matching methods specially designed for remote sensing image automatic registration was proposed based on analyzing of classical algorithms. An auto-registration framework under manually-supervision was presented and an application system was introduced at the end of this article. This article made contribution on the following factors:1. A new feature extraction method using affine parameter template was adopted. This method held an imaging simulation to generate several affine parameter templates .These templates contained affine transformation parameters between photos from some typical shooting examples and the orthophotos. The examples should be chosen correctly based upon prior knowledge. Compensating operations was adopted for the target image template by template and the SIFT features was then extracted from these transformed target images. Stability of this algorithm is compared to traditional methods under several interferences including affine transformation, partly barrier, tonal distortion, loss of contrast, and Gaussian noise.2. Two matching algorithms have been discussed. The first one built LOWE's similarity criterion from un-adjacent key points, because affine-templates SIFT method extracted too many features from a single space point and led to separating capacity reduction. The second one was a random sampling optimization algorithm. Density functions for sampling were created from feature distance matrix, which was the key component of this method. Optimization algorithm made it more stable to find right pairs from a lot of similar features. Stability is confirmed by experiments against matched points amount, time expending, noise proof feature, and contrast adaptability.3. A detailed description for peripheral techniques supporting feature extraction under huge image data size has been illustrated at the end of this article. Morton pyramid mid-ware built on GDAL kept efficiency for huge image access. Semi automatic registration would be a practical idea. Through manual multi-scale ROI delimiting, feature extraction for the whole image has been avoided. A more strict constraint condition could be established by manual scale offset measuring so that searching performance could be increased. A accuracy test was presented at the end of this chapter.