New Shape Descriptor And Its Applications For Image Matching

Image matching, not only an important but also a fundamental problem in computer vision, has wide applications in many fields such as three-dimensional reconstruction, robot navigation and image retrieval. The main task for image matching is to recognize the corresponding points in two or more images which contain the same object but may under different viewing conditions. Generally speaking, there are two types of image matching techniques, one based on the textures and the other on the features of the images. A projective transformation usually occurs when two images of an object from different viewpoints are acquired. In such a case, or even merely under a shearing transformation, image distortions are inevitable. As a consequence, neither the texture-based image matching techniques nor most of the feature-based ones can be directly utilized in a successful manner. Accordingly, it is significant to seek efficient and fast image matching strategies under these complicated transformations.In this paper, we first provide a brief state-of-the-art review of image matching techniques and introduce some preliminaries including the basic concept of the projective transformation. We then elaborate our two newly proposed shape descriptors and their applications for image matching under projective and affine transformations. The main contributions of this work are listed as follows:(i) We propose a new projective invariant, the characteristic number, and give its explicit representations over a line and a triangle. A prerequisite for constructing a characteristic number is that there exists a line set that constitutes a closed loop, with an equal number of points selected on each line.(ii) The intersections of the image and the line that obtained by connecting the sampled points on the convex hull of this image, are used to calculate the characteristic number on this line, i.e., the loop cross ratio, so that the shape descriptor of this image is constructed. We then employ the dynamic time warping algorithm (DTW) to design the image matching method for projective transformations,(iii) To accelerate the image matching procedure, the intersections of the image and the triangle which is constructed by connecting every three sampled points on the convex hull of this image, are obtained to calculate the triangular characteristic number over this triangle. Repeating such approach for all the possible triangles gives the shape descriptor of this image. The fast image matching strategy follows by combining this shape descriptor with histogram matching method for affine transformations,(iv) We analyze and compare the experimental results by employing some known classical image matching methods and our newly proposed ones for various scenes and images. The results show that our first approach outperform the others under projective transformations and our second one is faster and more appropriate for affine transformations especially for those containing shear transformations.