Title   name
KSIAM 2006 Annual Meeting
 
  Title
  Speaker   Hahn, Jooyoung  
  Date 2006-11-25
  Host
  Place 건국대학교
  File  의 1 번째 Real Media 동영상입니다. 의 1 번째 강연자료입니다.
 
Abstract : In the segmentation problems to extract objects from an image to make, for examples, 3D VR(virtual reality) contents or to estimate sizes of objects, a key issue is fine segmentation whichmeans that the objects can be extracted without visual loss of detailed shapes. Our research ismotivated by making 3D VR contents of commercial products. It makes an e-catalog that customerscan browse a product in three dimensional virtual space on internet markets. A commonway of making a 3D VR content starts from taking hundreds of photographs of a product withdifferent view angles in a photo studio. The most difficult step is to extract the product from abackground without visual loss of detailed shapes. The images taken in the studio have wellknowndifficulties in segmentation problems even though they usually have simple backgroundcolors and small amount of noises such as JPEG artifacts. These mainly come from lighting conditionsin the studio and complex shapes of products. Most of lighting conditions make shadowswhich cause weak boundaries between objects and the background. More serious weak boundariesare produced by a reflection on some parts of an object due to bright lighting conditionsand properties of materials of the object. It changes colors of objects into almost white which isnormally used as a background color. In addition, there are other difficulties; shapes of objectscan be highly non-convex.There have been a lot of boundary-based segmentation algorithms. The snake model in [1]has been a foundation of curve evolution based on the minimization of an energy. The geodesicactive contour model was introduced in [2] as the minimization of a weighted length. Althoughthe model has many advantages over the classical snake, it has drawbacks such as dependenceon positions of initial curves, incapacity for capturing weak boundaries when an image hasboth weak and strong boundaries, and slow convergence in non-convex boundaries. Numerousmodifications of the snake model and the geodesic active contour model have been developedto address these drawbacks. In [3], gradient vector flow was proposed for a fast convergenceto the non-convex boundaries. In [4], a curvature vector flow was introduced to overcome alimitation of [3] for segmenting highly non-convex shapes. In [5], the region-aided geometricsnake was proposed for more robust detection of weak edges. If an object in an image has bothweak boundaries and highly non-convex shapes, most of boundary-based segmentation algorithmssuffer from capturing such boundaries all around the object. Even though they capturethe boundaries, it is not enough to be a fine segmentation for extracting the detailed object froman image.In this talk, we propose a fine segmentation algorithm for extracting objects in an image,which have both weak boundaries and highly non-convex shapes. There are two main concepts, geometric attraction-driven flow (GADF) and edge-regions, which are combined to detectboundaries of objects in a sub-pixel resolution. Since an image is a two dimensional manifold,we obtain GADF by comparing two lengths of curves along the direction of the largest changein the manifold. Edge-regions contain most of edges. We compute inward fluxes in the gradientfield of a strength of edges to obtain such regions by which we construct initial curves close toboundaries of objects. Since the orientation of GADF near boundaries of objects points to edgesfrom each side of the boundaries regardless of strength of edges, we can segment the objects bysolving a simple advection equation of curves in the flow. It naturally solves problems of a slowconvergence near highly non-convex boundaries and a leakage over weak edges. According tothe purpose of segmentation, for examples, fine extraction of objects or measurement of sizesof objects, we additionally propose a local region competition algorithm to obtain perceptibleboundaries which are used for extraction of objects without visual loss of detailed shapes. Wehave successfully accomplished the segmentation of objects from images taken in the studio.Our algorithm can be applied to other kinds of segmentation problems by taking the appropriatestrategy for selecting the edge-regions. An example is to extract aphids from images of soybeanleaves. We may count the number of aphids that live on the sampled leaves and obtain an exactsize of each aphid.With those information, farmers can get the appropriate time to dust powder.