Symbol Sequence Recognition Using Viterbi Algorithm and Probability Approximation Table

Symbol sequence recognition performance can be improved by using preliminary information about symbols' order, types, etc. We propose an approach to increase the accuracy of symbol sequence recognition in automobile license number plates. It is based on the Viterbi algorithm, information about known sequence templates to estimate symbols' probabilities, thresholding of low-probability predictions and usage of the triplet loss during training of neural network detector. Results of experiments show that our approach increases the recognition accuracy and reduces the number of incorrectly recognized samples.