Manuscripts indexation using cognitive integration : from pixels to semantics - Comparative studies and enhancement of the state of the art – COGNILEGO

During the ANR blanche ANCL (2007-2009) with LSIS and LNIA, a project studying how the children learn to read, we demonstrated arguments in favor of a self-organized memorization of words (Dufau 2010). The model used in the study was a self-organizing map (Kohonen 1977) fed by inputs of letter bigrams (Grainger 2006). Given the strong performance of this model in orthographic reading, we want to use the same concept to explore word recognition.

Our concept suggests that the cortex does not process information; it memorizes information. This memorization preserves the topology and the density of the probability of the data to which the cortex is exposed. In our model, each level of memorization constitutes a hierarchical and functional level. Like S. Dehaene (2007), we estimate that ten levels are required to go from a pixilated image to a sequence of words.

The aim of the current project is to develop and test this word recognition model, called the Cogni-Lego. Our cognition-oriented proposals are original and state-of-the-art. After three years of research in the ANR blanche ANCL project, we propose to now integrate our findings with handwriting recognition models. The modular approach here proposed will allow us to complement the statistical approach of our industrial partner's handwriting recognition model. In the end, we aim to improve the overall performance of both our partner and our own models and to create a viable prototype of cognitive handwriting recognition.