In this paper, a colour text/graphics segmentation is proposed. Firstly, it takes advantage of colour properties by computing a relevant hybrid colour model. Then an edge detection is performed to construct a binary image composed of contour information. From this contour image, connected components are classified according to a graph representation. Text and graphic diversity is taken into account by a prototype selection scheme for structural data. Finally, the approach is evaluated on colour cadastral maps and a comparative study is presented.