Research On Video Based Handwritten Chinese Character Recognition

Video based handwritten Chinese character recognition is a kind of new development in the field of the handwritten Chinese character recognition (HCCR), which has the advantage of both off-line and online methods, but requires more for each operations involved. Around this task, three aspects of the problems are investigated in this paper, which include the feature extraction, the design of classifier & feature dimension reduction and video based handwritten Chinese character extraction &recognition system. The main work is as follows: 1. A new stroke-based directional decomposition approach for the handwritten Chinese characters is suggested. Considering the pixel distribution around the stroke pixel indicates the directional characteristic of the stroke, the directional characteristic number is defined at first, and then an algorithm for directional decomposition of the stroke is presented. Comparing with the traditional decomposition methods, it need not extracting the skeleton or contour, so can not only avoid the disadvantage of contour based methods being sensitive to different widths and distortions of the stroke, but also improve the blurring strokes and the losing of the important information of the stroke caused by skeleton based methods. 2. An elastic meshing feature extraction method based on the stroke density is proposed. Based on the stroke density definition used in the nonlinear shape normalization, a new kind of elastic meshing is constructed and applied in the feature extraction. The method can not only absorb the variations of strokes in different handwritings, but also avoid the unnatural and irregular width of the strokes that often occur in the nonlinear shape normalization. Meantime, its computation is less than that of the feature extraction method based on the nonlinear shape normalization. 3. By introducing the elastic meshes into the Gabor feature, an improved Gabor feature extraction method for the handwritten Chinese characters is present. The distribution of the sampling points can affect the performance of Gabor feature greatly. Based on the stroke distribution, the elastic meshes are constructed at first, and then the centers of the meshes are adopted as the sampling points. Comparing to the traditional methods that choose the sampling points uniformly, it can adapt variation in handwritings more effectively. 4. According to the characteristics of support vector machine (SVM), a SVM based approach for the large set of handwritten Chinese character recognition is suggested. As a new kind of machine learning algorithm, SVM is constructed based on the structure risk minimum (SRM), not only minimize the experiential risk, so the SVM classifier possesses the better generalization ability. However, for the large set of pattern recognition, the storage and the computation involved are too large either in the training or the recognizing phase. In order to solve the problem, a SVM based two-stage classification model for the handwritten Chinese characters is adopted. By the pre-classification of the minimum distance classifier, the training and the classification of SVM are confined to the candidate categories. Therefore, without reducing the classification ability, the number of the training samples, the training time and the support vector number decrease quite a lot for each of the binary classifiers. Meanwhile, the recognition velocity is also improved. And then, a weighted strategy for the classification is present, which improves the recognition rate effectively. 5. PCA(principal components analysis)&LDA(linear discriminant analysis) based feature dimension reduction for the handwritten Chinese characters is investigated at first, and then, a multi-channel PCA model based handwritten Chinese character recognition method is proposed. PCA&LDA can effectively reduce the dimension of the feature vectors without decreasing recognition rate markedly, while LDA can improve the performance of recognition with certain dimension reduction rates. In terms of the structure characteristic of the Chinese characters, the multi-channel PCA model based recognition approach decomposes a handwritten Chinese character into four directional sub-patterns at first, namely, horizontal, vertical, left up diagonal and right up diagonal sub-pattern, each of which could be modeled by its principal components. And then, a multi-channel PCA model for each category of the handwritten Chinese characters is constructed respectively, and the model's reconstruction error is used as a matching measure for the handwritten Chinese character recognition. Comparing to the direct reconstruction with PCA, the method can not only exploit principal components'ability for representing the handwritten Chinese characters, but also effectively reduce the training time for modeling. 6. A video-based online handwritten Chinese character recognition method is suggested and realized. Through a video camera, the handwriting on the standard paper is captured in real time at first. And then, the handwritten Chinese characters are extracted by applying image processing technique, and recognized online. Interms of the shape characteristic of both the handwritten stroke and the pen &hand, an updating algorithm for pentip template, a fast pentip matching algorithm and a mathematical morphology based handwritten Chinese character extraction approach are proposed respectively. Video based handwritten Chinese character recognition is a multi-discipline, comprehensive research item, which can realize a new kind of more natural video based approach for inputting the handwritten Chinese characters, and is far-reaching significance in theory and application.