Training an on-line handwriting recognizer

Abstract

Character model graphs are created, and the parameters of the model graphs are adjusted to optimize character recognition performed with the model graphs. In effect the character recognizer using the model graphs is trained. The model graphs are created in three stages. First, a vector quantization process is used on a set of raw samples of handwriting symbols to create a smaller set of generalized reference characters or symbols. Second, a character reference model graph structure is created by merging each generalized form model graph of the same character into a single character reference model graph. The merging is based on weighted Euclidian distance between parts of trajectory assigned to graph edges. As a last part of this second stage 'type-similarity' vectors are assigned to model edges to describe similarities of given model edge to each shape and to each possible quantized value of other input graph edge parameters. Thus, similarity functions, or similarity values, are defined by different tables on different model edges. In the third stage, model creation further consists of minimizing recognition error by adjusting model graphs parameters. An appropriate smoothing approximation is used in the calculation of similarity score between input graph and model graphs. The input graph represents a word from a work sample set used for training, i.e. adjusting the model graph parameters. A recognition error is calculated as a function of the difference between similarity scores for best answers and the one correct answer for the word being recognized. The gradient of the recognition error as a function of change in parameters is computed and used to adjust the parameters. Model graphs with adjusted parameters are then used to recognize the words in a test set, and a percent of correct recognitions in the test set is calculated. The recognition error calculation with the work set, the parameter adjustment process, and the calculation of the percent of correct recognitions with the test set is repeated. After a number of iterations of this process, the optimum set of parameters for the model graphs will be found.