Study On Aerial Image Registration

Image registration is a very important technique in remote sensing and still has many unsolved problems. Firstly, the ambiguity caused by dynamical objects, illumination change, large geometric transformation, similar patterns and low overlapping area between images can not be solved just by improving feature detection method. Accurate point matching is a critical and challenging process in feature-based image registration, especially for images with a monotonous background. Secondly, as there are mass data to be processed in remote sensing, the image registration algorithms are expected to be effective. To solve these problems, feature descriptor and feature matching are explored. Four improved algorithms are proposed as follows.1. Due to sea targets having no fixed shape and the mismatches in the matching result of Iterative Closest Point (ICP), an image registration algorithm using invariants-based similarity and improved ICP is proposed for registering images with dynamical objects. There are two stages in this algorithm:coarse registration and fine registration. Invariants-based similarity and relative spatial distance are applied to coarse registration. Then an improved ICP algorithm is used for registering images accurately by combining the ICP and a method of solving assignment problem to deal with mismatches. Compared with traditional ICP and NCC, the accuracy of the proposed algorithm is highly improved.2. It is difficult to find two identical regions in the images with low overlapping area and similar patterns. To tackle this problem, an image registration algorithm based on triangle regions is proposed. In this algorithm, relative moment affine invariants are used to evaluate the similarity of two triangle regions, then a new global feature matching method based on Genetic Algorithm is proposed to match the feature points accurately. Experimental results show that the algorithm works well to register images with low overlapping area and similar patterns.3. The intensity of an image pair may be different when they are taken with nonuniform lighting conditions or change in temperature, and MultiScale Autoconvolution (MSA) has difficulty in registering images with nonuniform lighting conditions, so an image registration algorithm based on an illumination and affine invariant is proposed. In this algorithm, an illumination and affine invariant called Illumination Invariant MultiScale Autoconvolution (IIMSA) is proposed to describe triangle regions and evaluate their similarity. IIMSA is the combination of MSA and MultiScale Retinex (MSR). Based on IIMSA, outliers are removed by a global matching strategy. Experiment results demonstrate that the algorithm is suitable for registering images with big illuminant changes.4. The efficiency and precision of traditional image registration algorithm can't meet the need of some system which requires high real-time, so a simple and robust feature point matching algorithm, called Restricted Spatial Order Constraints (RSOC), is proposed to improve the efficiency and accuracy. In RSOC, both local structure and global information are considered. Based on adjacent spatial order, an affine invariant descriptor is defined and point matching is formulated as an optimization problem. A graph matching method is used to solve it and yields two matched graphs with minimum global transformation error. In order to eliminate dubious matches, a filtering strategy is designed, which integrates two-way spatial order constraints and two decision criteria restrictions. Numerous experiments show that RSOC obtains the highest precision and stability.