Matching Of Images With Projective Distortion Using TILT

Image matching is to estimate corresponding relationship between two or more images of the same scene or objects in different times, from different sensors, or from different viewpoints. It plays a vital role in the field of computer vision and is the foundation of image registration, image fusion, object recognition and 3D reconstruction. Since 1970 s, studies on the algorithms of image matching have been drawn more and more attentions. Based on the differences of the transformation model between images, algorithms of image matching can be divided into three categories, i.e., similarity image matching, affine image matching and projective image matching. Traditional methods work effectively for images with less severe deformations such as similarity and affine transformation, but not for images with projective distortion. Matching of images with projective distortion becomes a difficulty for researchers. In this dissertation, a matching method for images with low-rank textures and projective distortion is proposed.Firstly, this dissertation makes a classification for the traditional image matching algorithms. These methods are mainly divided into three categories, i.e., intensity based, feature based and mathematic model based ones. Moreover, the feature based image matching methods, such as SIFT-based, ASIFT-based and MSER-based ones are described in detail.Secondly, the key idea of the transform invariant low-rank texture(TILT) is described. The feasibility of TILT in dealing with matching of images with projective distortion is discussed as well. Considering the characteristics of TILT in recovering intrinsic low-rank texture and the advantage of the traditional image matching method, an image matching method based on transform invariant low-rank texture(TILT) is proposed in this dissertation. This method accomplishes matching of images with projective distortion by three procedures:(1) An automatic low-rank texture region detection method is presented. With the proposed texture region detection method, the low-rank textures will be automatically selected from the original input images before they are fed into TILT.(2) A projective image matching method is proposed based on the TILT. Different from the traditional methods that try to directly seek local affine or projective invariant features from the input images, the proposed method reduces the problem of matching images with projective distortion to a problem of matching images with translation and scale distortions via TILT.(3) A feature based image matching method is employed to estimate the corresponding matched point pairs between the rectified images.Besides, A novel descriptor is introduced for each considered point when the two rectified images are matched. In addition to the local information from each considered point, the geometric shape information of the pixels around each considered point is employed in the proposed descriptor. This further improves the distinctiveness of each considered point to some extent, especially for those images with repetitive patterns or symmetric structures.At last, experiments are employed to demonstrate the validity of the proposed method. All the experiments are implemented with Matlab programming in the Windows7 environment. The results of the experiments demonstrate that the proposed method performs better than other existing methods in matching of images with projective distortion. More important, the introduction of the geometric shape information further improves the correct matching rate of the proposed method, especially for the images with many repetitive patterns or symmetric structures.