Arabic CWR Based on Correlation of Normalized Signatures of Words Images

摘要

The traditional methods for Arabic OCR (AOCR) based on segmentation of each word into a set of characters. The Arabic language is of cursive nature, and the character's shape depends on its position in the word. There are about 100 shape of the characters have to be classified, and some of them may be overlapped. Our approach use a normalized signature of the time signal of the pulse coupled neural network PCNN, supported with some shape primitives to represent the number of the word complementary and their positions within the image of the word. A lookup dictionary of words with its signatures was constructed, and structured in groups using a decision tree. The tested signature was routed through the tree to the nearest group, and then the signature and its related word with higher correlation within the selected group will be the classified. This method overcome many difficulties arise in cursive word recognition CWR for printed script with different font type and size; also it shows higher accuracy for the classification process, 96%.