Maximum-Minimum Similarity Training for Text Extraction

This paper proposes a maximum-minimum similarity training algorithm to optimize the parameters in the effective method of text extraction based on Gaussian mixture modeling of neighbor characters. The maximum-minimum similarity training (MMS) methods optimize recognizer performance through maximizing the similarities of positive samples and minimizing the similarities of negative samples. Based on this approach to discriminative training, we defined the objective function for text extraction, and used the gradient descent method to search the minimum of the objective function and the optimum parameters for the text extraction method. The experimental results of text extraction show the effectiveness of MMS training in text extraction. Compared with the maximum likelihood estimation of parameters from Expectation Maximization (EM) algorithm, the training results after MMS made the performance of text extraction improved greatly. The recall rate of 98.55% and the precision rate of 93.56% were achieved. The experimental results also show that the maximum-minimum similarity (MMS) training behaved better than the commonly used discriminative training of the minimum classification error (MCE).