Research On Key Technologies Of Dynamic Text Recognition

Text recognition (TR, text recognition) is an important part in the field of machine vision. It is an interdisciplinary subject which involving image processing, pattern recognition and data mining etc. Text recognition has been more and more widely applied in many different industries such as traffic, banking, post and office automation etc.The Text recognition system mainly consists of four parts: Text detection, single-character segmentation, feature extraction and character recognition. In these areas, the main questions focused on the accuracy of text detection, character segmentation and character recognition, and the merits of feature. Each step was studied in this paper and a dynamic text recognition system was given at last.In order to increase the reliability of detection, this paper improves the traditional algorithm which combines edge detection and morphological transformation. Prior knowledge and contour tracking are used to achieve precise text locating. In this model, we use rectangular contour angle to detect inclination angle of the rotating plate and correct it effectively. Line scanning method was used to determine the character's up and down scale. For character segmentation, in order to achieve real-time and high reliability, coarse segmentation and secondary segmentation method which based on prior knowledge were adopted.In feature extraction, we use three methods including coarse-grid method, 13-point feature extraction method and PCA method, and made a comparative analysis between coarse-grid method and 13-point feature extraction method. The experimental results demonstrate that coarse-grid method is better in divisibility and weaker in stability than 13-point feature extraction method. For character recognition, we use Template Matching classifier integrating SVM classifier. Template matching classifier is good at classifying small character set, while SVM has good generalization and adaptability. With the combination of those two classifiers, the performance of recognition system was improved.In this paper, the programming language is C++, and the compiler environment is VS2005 and OpenCV. Libsvm (SVM development tools) was induced into the system.System test shows that good recognition results were obtained by the system in the detection, segmentation and recognition tasks.