2000 Volume 14 Issue 2
Published: 17 April 2000
-
Select all|
-
Review2000, 14(2): 2-7,20.In this paper we describe the Testing System on Handwritten Chinese Character Recognition used by 863 High-Tech Program in 1998. We introduce the Testing Outline , Selection of Testing Samples and the Testing Result . Finally , we give some suggestions.
-
Review2000, 14(2): 8-13.This paper presents a methord for Chinese document layout analysis based on component . This methord mostly bases on a bottom-up approach ,it also benefits from a top-down approach and a concept of component . The concept of component lets the methord have a clear structure and reduces the times of scanning picture. Union the bottom-up approach and the top-down approach lets the methord have a high efficiency ,precision and adaptability. We use a two-dimensional orderly tree structure to organize document and comoponents. It improves the seaching speed and gives a convenience for application and document description.
-
Review2000, 14(2): 14-20."Imperial Collection of Four" is a sutra and representation of Chinese antient books. So the digitalization works of this Collection will accumulate and provide experiences for other antient books. This system is the pre-processing system of costumized OCR system for the digitized publication of "Imperial Collection of Four". The main function of this system is to analysis and undterstand the page images scanned from the Collection , then to seperate the Chinese characters in them for the use of recognition and statistics ,meanwhile extracting the layout structures for rebuilding and publishing. The design of the system adoptted top-down approaches with bottom-up ones ,and also adoptted automatic processings with manual correcting. In application , this system has been used to process a large numbers of page images ,and has shown efficient and satisfiable performance. It provides a stable ground for the pre-processing works ,and builds up a good situation for learning and recognition procedures of the recogintion system.
-
Review2000, 14(2): 21-25.In this paper ,we propose a Chinese business cards recognition system ,and analyze the structure of it . The first part of the system is the pretreatment of business card image. In this part ,we have to solve many practical problems. Then ,the system analyze the layout of the business card , divide it into several blocks. In follow ,the characters in every block are recognized ,and the result is further understood by using knowledge rules. After whole process ,the information of business card is imported to the computer automatically.
-
2000, 14(2): 26-30,48.In Chinese Character Recognition ,the classifier was designed as word classifier whose classification unit is a word in the past . The output of word classifier is always a set of candidate words that are similar with await-recognised words in structure of word. It is the primary reason that make mistakes in post-level recognition. To overcome disadvantage of word classifier , the strategy and method of phrase classifier designing whose classification unit is phrase are proposed. The experiments results prove that phrase classifier is superior to word classifier in rate and speed of classification.
-
Review2000, 14(2): 31-36,54.Recognition of multi-font characters becomes favorable research area in OCR by now. In this paper ,based on Support Vector techniques and SOFM neural network ,by use of a grorp of features that are complementary in their description of geometrical and topological structure of character ,we have proposed a recognition system. It include an Optimal Margin lingual classifier based on Support Vector techniques ,and a three-step Chinese character classifier based on growing self-organizing neural network. This system adopts both of statistical template matching and structure analysis methods. At the end ,following the experimental data ,a conclusion is made that this system is feasible and effective.
-
Review2000, 14(2): 37-42.In this paper ,We present a new handwritten digit recognition method based on wavelet networks and modular neural networks. Approximation ability of wavelet decomposition for general continuous functions and learning ability of neural networks can be combined to form the wavelet networks which has a nice feature descriptive ability ,and can be used as a feature extraction tools. The modular neural networks is used as a classifier , it reduce a k-class classification problems into a set of two-class classification problem. Using the modified BP algorithm which selects a suitable direction in each iteration of training phase ,the rate and accurate of convergence can be increased significantly. We test the method using U. S NIST character base , the result is good. The method can be also used to recognize the general plan objects.
-
Review2000, 14(2): 43-48.There are a lot of Chinese characters with many fonts in a topographical map. Some of them turn larger to surpass the scheduled threshold value or to lose their former structure traits because they are linked to other symbols ,which adds great difficulties to extraction. An efficient algorithm in search for the best segmentation points is presented in this paper to remove the adhesion and to extract characters. First ,fix a local searching area around the extracted characters. If there have large objects in the area , some characters may adhere to them potentially. Secondly , ramification points and extreme points in the images as the apexes ,lines as the sides ,set up corresponding graphics ;then find out the best segmentation points in the graphics to segment the symbols into several separate parts. At last ,extract the characters with the method of analyzing the structure of connected components.
-
Review2000, 14(2): 49-54.Form analysis is the first step in automatic form processing. This paper introduces a general form analysis method based on line extraction and completion ,which exploits thoroughly the property of form. We use a vectorization algorithm to extract form lines from run-length connect graph which is calculated first ,and in the same time the skew angle is detected. Lines are adjusted according to the characteristic of form. Then all critical points are calculated from which form cell description of the form can be derived based on some rules to complete the form lines.By applying this method ,broken-line and character-line-overlap problems can be solved and the structure of forms can be analyzed correctly.
-
Review2000, 14(2): 55-59.This paper studies the linear separability of Chinese characters through experiment and discusses the probability of applying linear classifier in Chinese character recognition. Two linear classifiers are used in the experiment . They are Fisher Linear Discriminant (FLD) and Perceptron. The experiment examines the linear separability of a pair of Chinese characters. The result is pretty satisfactory. Especially ,the result of Fisher Linear Discriminant is very good. There are only 4.25 millionths pairs that cannot be linear separated. This shows that the application of linear classifier in Chinese character recognition is potential. On the other hand ,this method can be applied to test the features used in recognition.
-
Review2000, 14(2): 60-64.A novel approach based on cascaded grouped BP network is proposed to classify handwritten digitals. According to the concept of simplify each classification task ,the number of classes for each net to classify is reduced to improve the accuracy. The whole system is divided into two cascades. The first step ,rough classification is to select the first two candidates. The second ,fine classification ,is to classify two specific kind of patterns. So the accuracy is improved because of the simplification of each classification task. Our system achieves the substitute rate of 0.05% with the rejection rate of 5% based on our own database with about 100000 characters.
