Affine Invariant Features Extraction For Scene Matching

With the urgent demand of precision strike weapons in our country, research work on scence matching has important significance for national defense. However the template and real-time image are gotton in different time and weather conditions, from different viewpoints, by different sensors, so there is a large distortion between them, and the traditional matcing methods based on gray-scale template can not position object accurately. Therefore extracting invariant features from the original and distorted images attracts more and more attention of researchers. In addition, affine-invariant features extraction is an important component of many computer vision tasks such as image registration, object tracking, 3D reconstruction, and object recognition.This paper studies methods of affine invariant features extraction in details, and the primary research productions are as follows:(1) A method of object recognition based on covariant matrix is presented, which is applicable for object with less texture, homogeneous gray level, and strong contrast with background. Firstly, segment the object from background with auto multi-threshold method. Then calculate the covariant matrix of object area and define an ellipse with it, by decomposing the covariant matrix, normalize the ellipse to a cirle and compute the invariant feature vector. Finally, compute the affine matrix between objects, and translate the template for further recognition. The descriptor computed by the presented method is full affine invariant, and insentitive to change in illumination. In addition, according to the basic theory proposed, methods of computing affine parameters and measuring object 2D attitude are given, which have high accuracy.(2) The method for extracting global invariant features based on MSA transform is studied, which is applicable for object in complicated background. Two improved methods against illumination change are proposed, which are MSA combined with direction code and MSA histogram invariant. Then two strategies for reducing computation load are discussed, which are blocking the image and discarding high frequency Fourier coefficients. Experimental results show that the improved methods are robust to illumination change, and the computation load is about 1/8 of that of the original algorithm.(3) The method for extracting local invariants based on SIFT is studied, which is applicable for scene in which object is partly occluded. An image may have many areas that are locally similar to each other, and these multiple locally similar regions produce ambiguities when matching local descriptors, which may result in mismatches. A method integrating global scope is presented to resolve ambiguities while allow for non-rigid shape transformations, which can enhance the discrimination of descriptors and reduce the numbers of mismatches. In addition, according to properties of affine transformation, a method for rejecting mismatches is given, which is effective without effecting correct matches.