Artificial Handwritten Character Generation

Introduction :

A new character distortion method for off-line handwritten character recognition has been deveopped. By allowing to artificially create new characters images from real ones, this method can be applied to increase the diversity of the database that is used for training a classifier, which can then results in a significant improvement of the generalisation ability of the recognition system. This approach is particularly suitable as the classifier is based on an artificial neural network, such as a multilayer perceptron, which is able to assimilate the distortions of the characters during its training phase. Then, no more character distortion is applied during the recognition phase, so that recognition time is not penalised at all in practice. Most existing distortion methods require a thinning and a vectorization process of the characters, which is difficult to perform in the case of off-line character recognition. As a matter of fact, the writing process is then unavailable. The distortion method that is proposed here is thus based only on the bidimensional image of the characters. Moreover, the generation of new characters images from an original one thanks to the use of this method only depends on a few number of parameters. This offers the advantage that the parameters of the distortion algorithm are very easy to set out, so as to produce characters images that are different enough from the original ones to bring new useful information, as well as to avoid creation of over-noisy images.

Principle :

The principle of this new distortion method is based on the fact that if different relative shifts are applied to each of the four corners of the image that represents a character, the whole character goes through a distortion. This result is obtained by applying the different relative shifts to the corners of a bidimensional sampling grid, which is then used to resample the character image. A gap of this method is that the relative positions of some typical features are never modified. So as to cure this problem, the widths of the sampling intervals also vary according to a geometrical sequence. This allows to dilate or to contract different areas of the image. The final parameters of the distortion algorithm are thus the amplitude and the direction of the shifts applied to the corners of the sampling grid, and the value of the geometrical ratios which define the sampling intervals. These parameters may vary randomly, but have to stay below a pre-defined threshold to avoid generation of distortions of too high amplitude. To avoid apparition of discontinuities in the lay-out of the character, a bidimensional low-pass filtering, followed by an erosion process in order to thin the resulting image, can be applied. However, it has appeared that this may involves a leakage of discriminant information. In spite of a higher computation time, it is thus preferable to process to an over-sampling of the original image, which leads to new characters of very better quality.

Results :

The tests that were carried out on handprinted digits extracted from the NIST3 database have shown that this method allows to reduce significantly the misclassification error rate: by training so a Multilayer Perceptron as a classifier, the recognition rate that was obtained on an independent test set has been increased from 97.0 % to 98.1 %. These tests have also shown that the best results are obtained when using systematically a maximal value for the parameters of the distortion algorithm. In practice, this allows to make sure that a sufficient distortion is always applied to the character. The random nature of the distortion process is then enclosed in the direction of the shift applied to each corner of the sampling grid, and in the direction (contraction or dilation) induced by the values of the geometrical ratios.

Back to the Technologies page.

Back to the Pattern Recognition Homepage.

Send coments to gosselin@tcts.fpms.ac.be.