Maximum-minimum similarity training for text extraction

Authors:
Hui Fu;Xiabi Liu;Yunde Jia
Affiliations:
School of Computer Science and Technology, Beijing Institute of Technology, Beijing, P.R. China;School of Computer Science and Technology, Beijing Institute of Technology, Beijing, P.R. China;School of Computer Science and Technology, Beijing Institute of Technology, Beijing, P.R. China
Venue:
ICONIP'06 Proceedings of the 13th international conference on Neural information processing - Volume Part III
Year:
2006

Citing 8
Cited 0

TextFinder: An Automatic System to Detect and Recognize Text In Images

IEEE Transactions on Pattern Analysis and Machine Intelligence
Robust speech recognition method based on discriminative environment feature extraction

Journal of Computer Science and Technology
Video OCR for Digital News Archive

CAIVD '98 Proceedings of the 1998 International Workshop on Content-Based Access of Image and Video Databases (CAIVD '98)
Minimum Classification Error Training for Handwritten Character Recognition

ICPR '02 Proceedings of the 16 th International Conference on Pattern Recognition (ICPR'02) Volume 1 - Volume 1
Locating Characters in Scene Images Using Frequency Features

ICPR '02 Proceedings of the 16 th International Conference on Pattern Recognition (ICPR'02) Volume 3 - Volume 3
ICDAR 2003 Robust Reading Competitions

ICDAR '03 Proceedings of the Seventh International Conference on Document Analysis and Recognition - Volume 2
Text Extraction from Web Images Based on A Split-and-Merge Segmentation Method Using Colour Perception

ICPR '04 Proceedings of the Pattern Recognition, 17th International Conference on (ICPR'04) Volume 2 - Volume 02
Learning to Detect Scene Text Using a Higher-Order MRF with Belief Propagation

CVPRW '04 Proceedings of the 2004 Conference on Computer Vision and Pattern Recognition Workshop (CVPRW'04) Volume 6 - Volume 06

Quantified Score

Hi-index	0.01

Visualization

Abstract

In this paper, the discriminative training criterion of maximumminimum similarity (MMS) is used to improve the performance of text extraction based on Gaussian mixture modeling of neighbor characters. A recognizer is optimized in the MMS training through maximizing the similarities between observations and models from the same classes, and minimizing those for different classes. Based on this idea, we define the corresponding objective function for text extraction. Through minimizing the objective function by using the gradient descent method, the optimum parameters of our text extraction method are obtained. Compared with the maximum likelihood estimation (MLE) of parameters, the result trained with the MMS method makes the overall performance of text extraction improved greatly. The precision rate decreased little from 94.59% to 93.56%, but the recall rate increased a lot from 80.39% to 98.55%.