A structural representation for understanding line-drawing images

Abstract.

In this paper, we are concerned with the problem of finding a good and homogeneous representation to encode line-drawing documents (which may be handwritten). We propose a method in which the problems induced by a first-step skeletonization have been avoided. First, we vectorize the image, to get a fine description of the drawing, using only vectors and quadrilateral primitives. A structural graph is built with the primitives extracted from the initial line-drawing image. The objective is to manage attributes relative to elementary objects so as to provide a description of the spatial relationships (inclusion, junction, intersection, etc.) that exist between the graphics in the images. This is done with a representation that provides a global vision of the drawings. The capacity of the representation to evolve and to carry highly semantic information is also highlighted. Finally, we show how an architecture using this structural representation and a mechanism of perceptive cycles can lead to a high-quality interpretation of line drawings.