| The document image analysis has been intensively studied in the last decades, while Hidden Markov Model (HMM) turns out to be the mainstream method in image processing and computer vision. In this paper we strive to implement HMM in the most critical two steps of document image analysis: Binarization and Optical Character Recognition (OCR). We propose a HMM based binarization method, whose efficiency is demonstrated through the simulation results of document images with different degradations. At the same time we also introduce an edge based binarization method, which has lower computational burden and higher performance than other similar methods. Test results for OCR applications show 77% correct rate for HMM based binarization method and 67.3% for the edge based binarization method, while the best performance for other reference techniques is 57%.; The conventional HMM based classifier is a causal system, which means the deduction of the hidden states is unidirectionally obtained along a single path. Thereby, the performance of the HMM classifier can be degraded by the noise mixed in the signals, which is inevitable for the real world applications. Here we propose a new non-causal Self-Adaptive Hidden Markov Model (SAHMM) for Optical Character Recognition (OCR) application. The experiment results prove that it has stronger immunity to the noise than the conventional methods. We also extend the 1-D SAHMM into 2-D, where some new feature extraction methods and new architecture for the nodes in the model are introduced. The proposed 2-D SAHMM OCR, engine achieves recognition rates of up to 96% on the MNIST database, which is higher than any reported single HMM based OCR engine. |
