An Improved Neural Network Method For Segmentation Of Merged Characters

Computer-aided document-handling systems have been widely used, and OCR systems have found its place in OA, Quick Input, and etc. There are many mathematical expressions in science and technology literature and these expressions consist of special characters with complicated structure. The question how to use OCR for recognizing and understanding mathematical expressions in printed document is being studied. To this end, our research group has designed a recognition system for mathematical expressions in printed document. The system has the ability to recognize mathematical expressions in scanned files and to reconstruct the recognized expressions into a particular format such as LATEX or WORD. Based on the previous works in this respect, we in this thesis impose an improved neural network method for segmentation of merged characters so as to improve the performance of the recognition system.After a brief introduction of the construction of the system, this thesis mainly studies two problems: merged-character segmentation and character recognition.Merged Character Segmentation: Due to the quality of printer, cleanliness of paper, resolution of scanner, binarization and etc., symbols in scanned document may be merged and can not be easily recognized. In this thesis, an algorithm for segmentation of merged characters based on "Shortest Path" and "Modified SOFM" is presented. This proposed segmentation method behaves particularly good for poorly printed characters, and it can make a good supplement of the most often used shortest path method.Character Recognition: Recognizer is a key part in the system. Due to its good performance in pattern recognition field, artificial neural network has already been used in this part. First, a SOFM network works as rough-classifier, which classify similar symbols into the same group. Then, BP network are used as fine-classifiers, which identify symbols within a group. In this thesis, some explanations in terms of statistical learning are given for the output representation of BP neural networks, hoping that it will improve the performance of BP neural networks when they deal with classification problems.