Tham khảo tài liệu 'mobile robots perception & navigation part 5', kỹ thuật - công nghệ, cơ khí - chế tạo máy phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | Accurate Color Classification and Segmentation for Mobile Robots 151 Color image segmentation Color image segmentation can be defined as the process of decomposing an image into regions with certain meaning according to the contents and the application for that specific image Watt Policarpo 1998 . Here a region is a set of connected pixels sharing a set of attributes. Classic segmentation techniques can be divided in global approaches region-based approaches and edge-based approaches. Global or threshold approaches rely in the knowledge of pixel attributes. Thus it is required to provide with a set of attributes that bind the classes that must be identified within an image. In the case of color images this approach uses the color space as a 3D domain over which a set of groups or clusters will be identified for each color class. This technique is simple and efficient but depends heavily on a good threshold definition. Region-based approach consists in dividing an image in a set of regions that present similar properties. Techniques using this approach usually start by growing a region from an initial pixel or seed and expanding this region according to a set of homogeneity criteria. This approach presents two general problems. First an initial seed should be picked and this requires an additional process. And second it is usually hard to define and parameterize the homogeneity criteria. An incorrectly defined homogeneity criterion may lead to flooding problems where regions grow over the visible boundaries of the region or to prematurely stop the growth process. The edge-based approach uses edge detection to find a closed boundary that defines what lies inside and outside a region. The hypothesis in which this approach relies is that pixels in the boundary between two regions should be considerably different regarding properties such as color or intensity. However problems are produced in blurry areas of an image where colors are not very contrasting. In .