Cursive Word Recognition Using a Novel Feature Extraction Method and a Neural Network

In this paper, we present a holistic system for the recognition of cursive handwriting that utilizes a novel feature extraction method and a neural network. The Hough transform is a global line detection technique with the ability of extracting directional information presenting good tolerance to disconnections and noise and a moderate tolerance to distortion. However, its global nature is also its weakness because it does not capture well the information contained in the body of cursive words. For this reason, we have modified and reformulated this method as the correlation of the image with template line segments which is able to extract more local features while preserving the advantages of the previous method. In fact the new method is more general in the sense that it can also be used to detect complex features such as closed loops, cavities and to estimate the average stroke width. An important result is that this new method is also equivalent to a sigma-pi network closely connected with the well known feed-forward model of orientation selectivity [5]. Results obtained on a task of cursive word recognition indicate that the new method clearly outperforms the old method.