HMM-based off-line English cursive word recognition, which advances the development of both HMM theory and its applications, is still the highlight in the OCR field. On the other hand, the recognition techniques have many applications such as in the field of the postal address recognition, bank check reading, generic content recognition and so on. For the two reasons above, English cursive word recognition has been the focus of researchers all over the world. Firstly, we introduced and analyzed the applications of HMM in the off-line cursive English word recognition. Secondly, based on the introduction and the analysis, we designed and realized an HMM-based recognition system, which is a two-pass system recognition plus verification. The main contributions of this paper are as follow: 1、 In the step of preprocessing, the technology of image binary, noise removing, normalization and reference line detection are used. Since it is difficult to find reference line correctly depending on profile only, we integrate transition with profile. As a result, we can relatively detect the correct reference. 2、In the thesis, we used two groups of features. Each feature group is taken by means of sliding window. Although the width of sliding window is generally fixed or takes average value by current samples, the writing style always affects a lot. So we dynamic the width and overcome the effect of writing style in a certain extent. 3、 In the step of HMM-based recognition, we used implicit-based segment method. Word model is made of character model by linear connecting. For the difference of each character, the count of state in each model is unequal. 4、 The system we designed in this paper adopted two-pass structure. The first pass is HMM-based recognition. As the process of recognition is going on. the best segments of the top three candidates are traced through Viterbi algorithm. Then the result is sent to the second pass. The verification is performed through matching with reference points found by LVQ. At last we integrate these two results and get the final recognition output. To verify the system of this paper, two tests are taken in the samples of NIST and Cambridge respectively. The first sample is written by 500 persons, the second is written only by 1 person. The result is essentially satisfying.
修改评论