We describe a method for hand-drawn symbol recognition based on deformable template matching able to handle uncertainty and imprecision inherent to hand-drawing. Symbols are represented as a set of straight lines and their deformations as geometric transformations of these lines. Matching, however, is done over the original binary image to avoid loss of information during line detection. It is defined as an energy minimization problem, using a Bayesian framework which allows to combine fidelity to ideal shape of the symbol and flexibility to modify the symbol in order to get the best fit to the binary input image. Prior to matching, we find the best global transformation of the symbol to start the recognition process, based on the distance between symbol lines and image lines. We have applied this method to the recognition of dimensions and symbols in architectural floor plans and we show its flexibility to recognize distorted symbols. © Springer-Verlag Berlin Heidelberg 2000.
|Journal||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|Publication status||Published - 1 Jan 2000|