Two-dimensional PHMM for handwriting digits recognition

We used two-dimensional pseudo hidden Markov models (2D PHMM) for handwriting digits recognition. We used a 2D pseudo Viterbi algorithm for training the model. This algorithm compared to the traditional forward-backward algorithm had some advantages as follow. (1) It can get rig of the down overflow that may happen in the forward-backward algorithm. (2) As the algorithm used log calculation the accuracy of the calculation was much higher than that of the forward-backward algorithm. (3) The speed of calculation was much faster than the forward-backward algorithm. The database of handwriting digits used was collected at the Beijing Postal Center, the digits were scanned from letters zip code; of the 4000 handwriting digits collected, 2000 were used for the training set, and 2000 were used for the testing set. We used the pixel value as the recognition feature. The test result showed, the recognition rate of training set was 95%. The recognition rate of testing set was 92.6%. We used the horizontal and vertical directions as a super plane to build two types models at same time, and at the recognition phase if the result of the two models are different as a rejection condition, the recognition rate of training set is 98%, the recognition rate of testing set is 96%. If we used some feature which is rotate and move invariant the recognition rate may get higher