Hand-drawn symbol recognition in graphic documents using deformable template matching and a Bayesian framework

Hand-drawn symbols can take many different and distorted shapes from their ideal representation. Then, very flexible methods are needed to be able to handle unconstrained drawings. We propose to extend our previous work in hand-drawn symbol recognition based on a Bayesian framework and deformable template matching. This approach has enough flexibility to fit distorted shapes in the drawing while keeping fidelity to the ideal shape of the symbol. We define the similarity measure between an image and a symbol based on the distance from every pixel in the image to the lines in the symbol. Matching is carried out using an implementation of the EM algorithm. Thus, we can improve recognition rates and computation time with respect to our previous formulation based on a simulated annealing algorithm