Arabic optical character recognition method using hidden markov models and decision trees
View Patent ↗Disclosed is an Arabic optical character recognition method using Hidden Markov Models and decision trees, comprising: receiving an input image containing Arabic text, removing all diacritics from the input image by detecting a bounding box of each diacritic and comparing coordinates thereof to those of a bounding box of a text body, segmenting the input image into four layers, and conducting feature extraction on the segmented four layers, inputting results of feature extraction into a Hidden Markov Model thereby generating HMM models for representing each Arabic character, conducting iterative training of the HMM models until an overall likelihood criterion is satisfied, and inputting results of iterative training into a decision tree thereby predicting locations and the classes of the diacritics and producing final recognition results. The invention is capable of facilitating simple recognition of Arabic by utilizing writing feature thereof, and meanwhile featuring comparatively high recognition precision.
1. A method for establishing a HMM-based encoding network using connectivity rules of Arabic optical characters in a recognition process, the method comprising:
providing three non-emitting states in the HMM-based encoding network: a beginning non-emitting state disposed at the beginning of said network, an end non-emitting state disposed at the end of said network, and a middle non-emitting state disposed at the middle of said network;
connecting said beginning non-emitting state to an isolated position character hidden Markov model, and said isolated position character hidden Markov model to said end non-emitting state;
connecting said beginning non-emitting state to a beginning position character hidden Markov model, and said beginning position character hidden Markov model to said middle non-emitting state;
connecting said middle non-emitting state to an end position character hidden Markov model, and said end position character hidden Markov model to said end non-emitting state;
connecting in parallel said middle non-emitting state to a middle position character hidden Markov model, and connecting said middle position character hidden Markov model back to said middle non-emitting state thereby forming a loop between said middle non-emitting state and said middle position character hidden Markov model;
connecting in parallel said middle non-emitting state to an elongation hidden Markov model, and connecting said elongation hidden Markov model back to said middle non-emitting state thereby forming a loop between said middle non-emitting state and said elongation hidden Markov model; and
connecting said end non-emitting state to said beginning non-emitting state.
2. The method of claim 1 , further comprising:
providing an input for said HMM-based encoding network, the input comprising feature vectors; and
using said HMM-based encoding network and a Viterbi decoder to output the most optimal character sequence.
3. The method of claim 2 , further comprising combining an output of said HMM-based encoding network and Viterbi decoder with a diacritics feature using a decision tree, thereby producing a final recognition result.