Creating Quadrilateral Mosaics from Image Topographic Features Kazuyo Kojima ∗ 1 Shigeo Takahashi † The University of Tokyo Introduction 3 Image mosaicing is a method for partitioning an input container image into small image tiles. Human-crafted image mosaics usually respect not only local image features such as object silhouettes but also the global features such as image gradation. This is also justified because high-level human vision perceive rather global image gradation while low-level one identifies local image edges [M. J. Jones and Poggio 1997]. All conventional image mosaicing methods take a bottom-up approach in that they start with handling local features and then stitch them together to form a whole image mosaic [Elber and Wolberg 2003]. This bottom-up approach usually suffers from the lack of such local features as well as high-frequency noise because it relies solely on the local features. In this paper, we propose a top-down approach that first generates a mosaic respecting global features such as image gradations, and then incorporates local features into the underlying partition based on the global features. This approach allows us to generate a sound partitioning of the input image although the image contains unclear object silhouettes and high-frequency noise. Moreover, the method employs quadrilateral tiles as its mosaicing primitives because our visual system can easily perceive the iso-contour and gradient flows inherent in the input image by tracking their opposing edges. 2 Tomoyuki Nishita ‡ Results We have implemented our prototype system on a standard PC with Intel Core Duo 2.0GHz using the C++ language and OpenGL library. Figure 2 shows experimental results obtained by using our method. Each row of Figure 2 from left to right shows an original image, a mosaic only with global features, and a mosaic with both global and local features. Figure 2(a) shows a mosaic of a duck image. As seen in this figure, the present method can embed local features such as silhouettes of the duck without disturbing the tile arrangement respecting global image gradations. Figure 2(b) shows a mosaic generated from an image of a sunset scene. As shown in this figure, the present method successfully respects both the global and local image features such as color gradation and its profiles around the sun. Figure 2(c) shows a mosaic of a bird image. This figure reveals the robustness of the present method because, with the resultant tile arrangement, it can successfully discriminate the focused region of the bird from the unfocused background region on the input image. On the other hand, conventional methods cannot fully handle the blurred silhouettes of the bird together with the background image gradation. Method (a) Figure 1 shows steps in the present method. We first construct a quadrilateral mosaic that respects global features such as image gradations. We then incorporate local features such as object silhouettes specified by users into the existing global mosaic by referring to the dual network of the mosaic partition. Recoloring each mosaic tile with its average color allows us to finalize this image mosaic generation. Input image Output image global features local features quadrilateral partition respecting global feature Figure 1: Steps in the proposed method ∗ e-mail: [email protected] [email protected] ‡ e-mail: [email protected] † e-mail: (b) (c) Figure 2: Mosaic examples: (a) a duck, (b) sunset, and (c) a bird. 4 Future Work We plan to extend the present method to transform the quadrilateral mosaic into a mosaic with polygons of arbitrary shapes for simulating stained glass tiling, and to adjust the color tone of each tile in the mosaic by taking account of artistic styles and visual perception. References E LBER , G., AND W OLBERG , G. 2003. Rendering traditional mosaics. The Visual Computer 19, 1, 67–78. M. J. J ONES , P. S INHA , T. V., AND P OGGIO, T. 1997. Top-down learning of low-level vision tasks. Current Biology 7, 12, 991– 994.
© Copyright 2026 Paperzz