Hybrid Model of Post-Processing Techniques for Arabic OCR

OCR is used to extract text contained in an image. One of the stages in OCR is the post-processing and it corrects the errors of OCR output text. The OCR multiple outputs approach consists of three processes: differentiation, alignment, and voting. Existing differentiation techniques suffer from the loss of important features as it uses N-versions of input images. On the other hand, alignment techniques in the literature are based on approximation while the voting process is not context-aware. These drawbacks lead to a high error rate in OCR. This research proposed three improved techniques of differentiation, alignment, and voting to overcome the identified drawbacks. These techniques were later combined into a hybrid model that can recognize the characters of the Arabic language. Each of the proposed technique was separately evaluated against three other relevant existing techniques. Experimental results showed a relative decrease in error rate on all measurements for the evaluated techniques. Similarly, the hybrid model also obtained lower WER, CER, and NWER by 30.35%, 52.42%, and 47.86% respectively when compared to the three relevant existing models.