| The adaptive recognition techniques and an adaptive character recognition system for multiple small-character-sets are introduced in this paper. The background of this project is to develop the OCR Based Automated Entry and Management System of Maintaining Cards of Train Wheels for a certain railway unit. The characters which are to be recognized include the uppercase English letters, digits, some special symbols and some Chinese characters, and the character-set may be changed in the future. The required recognition rate is fairly high. On the basis of the deep research of the OCR techniques, we accomplished the research project successfully.We researched the principle of the application of various neural network models in the character recognition, and by compare, we think that the MLPs are suitable for using in our system. On the other hand, the learning sub-space method (LSM) is a powerful tool to smartly construct characters classifiers as neural networks. We researched the principle of the LSM, put forward an enhanced rejecting rule of LSM classifiers, and improved on the basic LSM algorithm and the Oja's dimension selection algorithm of subspace.One of the keys to implement the adaptive character recognition system is to extract right features adaptively. The main factor, which affects the adaptability of OCR systems, is that human do too much work in the feature extraction and selection. Our point is to automatically generate the feature extraction schemes with few human's experiences. We do the research in two directions.One is to find the invariant features through learning. We discussed some statistical feature extraction methods, and the application of ASSOM neural network in feature extraction.The other direction of our research is to select the features, which are the most suitable for classifying the specific character-set from large amount of candidate features. We introduced the feature selection method based on rough set theory, and put forward a rapid property reductionalgorithm of rough set. After the roughly feature selection, the dimension of the feature vector is further decreased by PCA method, and the correlations between the features are eliminated.In addition to the recognition method based on machine learning, we researched the character recognition method base on deformable templates too. The deformable template methods (DTM) can resolve some problems that statistical and ANN recognition methods can not accomplish. The prominent advantage of DTMs is that they can make full use of the prior knowledge on the shape of the characters, and can deal with deformed characters without training work. It's much valuable for recognition small character-set. We improved the methods of Michael Revow and Kwok-Wai Cheung, and applied the method to multi-stroke characters such as English letters.In the research of the applied OCR system, we divided the large character-set into a serial of smaller character-sets by applying an intelligent form description and analysis method. This method defines the content of the forms before processing, every filling-in position of the characters can only be filled in only with few characters, so the large character-set can be divided according to this situation. It's a useful method to solve many practical problems. In addition, we produced a set of intelligent layout analysis and segmentation algorithm of forms based on the description method of forms, including a segmentation algorithm of single color (black) forms.We can't expect to get a satisfactory recognition rate by using a uniform recognition program to recognize these character-sets. We must modify and optimize the recognition programs according to the character-sets. We developed an adaptive character recognition system which can select optimal recognition schemes according to the target character-set. There is an adaptive controlling module in this system, it can automatically select most suitable recognition scheme through the result of recognition tests. The main idea is as fol...
|
PDF Full Text Request