Printing knowledge

Talking about the printed text recognition in research methods

Structural pattern recognition is the main method in the study of early Chinese character recognition. Its starting point is the composition and structure of Chinese characters. Speaking from the form of the Chinese characters, Chinese characters by stroke (point anyway whets), composed of radicals; can also think that the Chinese character is made up of smaller structural motifs. By the structure of primitive characters and their relationships can be accurately be described, as an article by Word, words, phrases and sentences in the grammar rule consists. So this method is also called syntactic pattern recognition. Identification, information and syntax analysis methods for identifying the structure, similar to a logical reasoner.
statistical pattern recognition of Chinese character is the character dot matrix as a whole, its features from this after a lot of statistics are as a whole. Statistical characteristics are characterized by strong anti-interference, matching and sorting algorithm is simple and easy to implement. Downside is subdivision was weak and poor ability to distinguish between similar words. Common methods of statistical pattern recognition are:
(1), use the transform method. Binary to character image transform (Walsh, Hardama transform) or more complex transformations (such as Karhunen-Loeve, Fourier,Cosine,Slant transformation, and so on), the transformed feature greatly reduces the number of dimensions. However, these transformations are not rotationally invariant, so the inclination deformation of character recognition will have larger deviations. Binary transformation computation is simple, but transform features no clear physical meaning. K-l transformation from the minimum mean square error of angle is the best, but the operation is too large, to practical. In short, features high computational complexity, and some weaknesses.
(2) template. Template matching does not require feature extraction process. Image of the character directly as a feature, compared with the dictionary template, template class that is most similar to the recognition result. This method is simple and easy, and can be processed in parallel, but can only identify a template of the same size, the same font characters, tilt, both thick and thin strokes without a good ability to adapt.
Business scope
Contacts us

Phone: 0871-8672252

Fax: 0871-8672252