The hottest printed Chinese character recognition

  • Detail

Printed Chinese character recognition system (III)

4.3 combination of statistical recognition and structural recognition

structural pattern recognition and statistical pattern recognition have their own advantages and disadvantages. With the deepening of our understanding of the two methods, the two methods are gradually integrating. The lattice feature is the product of this combination. The character image is evenly or unevenly divided into several areas, called "lattice". Look for various features in each grid, such as the proportion of stroke points to background points, the number of intersection points, stroke endpoints, the length of refined strokes, the stroke density of grid parts, and so on. The statistics of features are based on lattice. Even if there is an error in the statistics of individual points, it will not cause a great impact, which enhances the anti-interference of features. This method is being widely used

4.4 artificial neural network

artificial neural network (ANN) is a network structure that simulates human brain neuron cells. It is an adaptive nonlinear dynamic system composed of a large number of simple basic elements neurons interconnected. Although the current research on human brain neurons is still far from perfect, and we cannot be sure whether the working mode of ANN is the same as that of human brain neurons, Ann is attracting more and more attention

the structure and function of each neuron in Ann are relatively simple, but the combination of a large number of simple neurons can be very complex, so we can complete the complex functions of classification and recognition by adjusting the connection coefficient between neurons. Ann also has a certain adaptive learning and organization ability. Each "cell" of the network can work in parallel, and can complete complex functions such as classification and recognition by adjusting the connection coefficient between "cells". This is what von Neumann's computer can't do

ann can be used as a simple classifier (excluding feature extraction and selection), or as a fully functional classifier. In the classification problems with a small number of categories, such as the recognition of English letters and numbers, the image lattice of characters is often used as the input of neural networks directly. Unlike traditional pattern recognition methods, in this case, the features "extracted" by neural networks have no obvious physical meaning, but are stored in the connections of various neurons in neurophysics, eliminating the need for people to determine the method and implementation process of feature extraction. In this sense, Ann provides a possibility of "automatic character recognition". In addition, ANN classifier is a nonlinear classifier, which can provide a complex interface between classes that we can hardly imagine, which also provides a possible solution to complex classification problems

at present, the scale of ANN will be very large and the structure will be very complex for super multi class classification problems such as Chinese character recognition, which is far from practical. There are many reasons. The main reason is that we have not found perfect answers to many questions about the working mode of human brain and Ann itself

v. the latest development of Chinese character recognition technology

the most important index of Chinese character recognition is the recognition accuracy. The latest technology includes two aspects: first, using the comprehensive recognition method of combining optimized features to improve the accuracy and adaptability; Second, Chinese English bilingual mixed row recognition when the ratio of English to numbers exceeds 1/3

5.1 combination optimization special engineering plastics 1 generally refers to a comprehensive recognition method that can be used as structural materials to bear mechanical stress signs

extracting a single kind of features for Chinese character recognition, which is difficult to reduce the error rate and improve the anti-interference performance. Because the amount of information of Chinese characters used in this way is limited, it can not fully reflect the characteristics of Chinese characters. For any feature, there must be a "dead corner" of its recognition, that is, Chinese characters that are difficult to distinguish by using this feature. From the perspective of pattern recognition, if the space composed of all vectorized features of Chinese characters is called space (i=1,2,......), then using the information of the whole space Ω for Chinese character recognition will greatly enhance the anti-interference ability because the Chinese character information provided is very sufficient

however, in practical applications, the tradeoff between recognition accuracy, recognition speed (Computation) and system resources must be considered. Therefore, any practical OCR system only uses the information of some subspaces. Due to the defects of information, it is inevitable to encounter the problem of identifying "dead corners". The basic idea of "combination optimization feature method" to recognize Chinese characters is: first, on the basis of long-term research on Chinese character recognition, select a variety of statistical features based on Chinese character stroke structure, such as structural elements, which have good classification ability of intra class aggregation and inter class divergence; Secondly, a variety of Chinese character features complement each other organically, which greatly reduces the "dead corner" of Chinese character recognition, thereby improving the recognition rate

the comprehensive recognition method of "combined optimization features" is based on a full understanding of various methods, which is a knowledge-based recognition method, because it is targeted, gives full play to the advantages of each scheme, achieves a high recognition rate, and improves the operation efficiency of the system

5.2 Chinese English bilingual mixed row recognition

with the opening and development of the information industry, more and more English words appear in China's printed text materials. Especially in scientific and technological literature and publications, it is common that the proportion of English and figures often exceeds 1/3. English letters appear in the text line, and their size and height are very similar to the radicals in Chinese characters, so it is difficult to distinguish whether they are Chinese characters' radicals or English letters; The distance between letters in English words is different, and it is quite common to use S-type sensors for small force values of tension machines on the market at present; Chinese characters are based on horizontal and vertical strokes, while English is based on curves. Therefore, the key of Chinese English bilingual mixed row recognition lies in the correct discrimination and segmentation of Chinese characters and English letters. The traditional method of segmentation is to use the information of "height and position", but because there are many left and right separable characters in Chinese characters, their parts, regardless of height and width, are very close to English letters, such as "namely", "old" and so on; Moreover, the adhesion of English letters cannot be solved; In addition, many English two letter combinations and three letter combinations are printed by using a font that cannot be separated from the robot system by intelligent manufacturing, such as "fi". Therefore, on the basis of the criterion of "height and position", according to the "try error try" criterion, the means of "segmentation and secondary segmentation of Dong" is added. That is, for all possible segmentation situations, pre identify, and select the combination with the least error and the most logical habit of language

for English letters that are stuck together, there may be multiple letters that are continuously stuck together, so there are many combinations, and the types of adhesions are also quite different. The "exhaustive" trial segmentation takes too much time. Therefore, the method of "breaking up the whole into parts" is adopted. According to the best neighborhood search principle, the letter string is divided into two sub strings from the most reliable place based on the quality of the hydraulic oil and the projection information of the letter string in the horizontal and vertical directions; Then repeat the above steps in the two substrings until the length of the substring is about the average width of three English letters; Finally, the "exhaustive" trial segmentation is carried out, which greatly shortens the time of segmentation. Thus, the recognition of Chinese English bilingual mixed text is effectively solved

5.3 other key technologies of high-performance practical Chinese character recognition system

other key technologies of practical Chinese character recognition system mainly include:

(1) scanner automatic brightness adjustment (abj automatic brightness adjustment) technology

(2) neighborhood analysis technology for automatic input of printed forms

(3) ala automatic layout analysis technology

to sum up, the latest printed Chinese character recognition technology process is shown in Figure 3

Figure 3 latest printed Chinese character recognition technology process

Copyright © 2011 JIN SHI