Recognition of handwritten connected numerals based on dual cooperative neural network

Recognition of unconstrained, handwritten connected numerals based on a dual cooperative neural network (DCN) is presented. First, a sequence of connected numerals is segmented into regions of interest by a group of windows moving along the horizontal axis of numeral sequence. The windows of the group are distributed in position in such a way as to achieve transition invariance. Since DCN combines both Cartesian and log-polar representation of numerals, robustness to transition as well as to rotational and scaling variations can be achieved. Introduced also are the multiple matching schemes from different types of correlation between the input and templates, to handle multiple numerals overlapped or captured by a window. A two-stage supervised self-organization process is implemented for the automatic generation of templates of each numeral. A set of templates thus generated provides robustness to pattern variations due to distortions. A multilayer backpropagation network generates outputs with its trained weights based on a sequence of multiple matching scores from individual numerals. An experimental result is shown.