How to overcome some segmentation problems in a constrained handwritten Arabic character recognition system

When using usual segmentation algorithms we have problems in segmenting some handwritten characters like the letters `u´ and `w´ in Script Latin and U and U in Arabic. To overcome these problems we decided to correct some segmentation problems after the rejection or the ambiguousness decision of the recognition system for certain entries. Indeed, the major rejection problems are coming from errors in segmentation. The problem being the over segmentation inside the character, we decided to paste the segmented parts to rebuild the whole character form. Instead of pasting the image´s parts we `concatenated´ the characteristic vectors. First, we used an MLP in the recognition process but in doing so we couldn´t really estimate the improvement achieved by the `concatenation´ in the recognition rate. So, we replaced the MLP by a linear feedforward network and we performed the concatenation until the neuron´s activation reaches the best outcome. Satisfying results have been achieved.