A new method is proposed for feature extraction and representation of character and figure patterns using a vector field. The vector field is constructed from a distance transformation like the gradient vector of the electric field caused by a charged pattern. Features are defined by a set of the source points, sink points, and locally homogeneous domains of the vector field. It is shown that these features are effective for detecting holes (loops) and directional segments of the input pattern. The pattern and the frame can be segmented naturally and the segmentation represents the juxtapositional relation among the locally homogeneous domains of flow-out vectors from the pattern. The authors discuss the relation between the distance transformation that the vector field is based on and scale-space filtering which can be regarded as a kind of diffusion and/or blurring processing of the shape of pattern.
展开▼