Research On Key Technologies In Handwritten Numeral Recognition

Nowadays, Optical Character Recognition (OCR) plays a more important role in our society. As an important branch of the OCR technology, handwritten numeral recognition can be applied to many applications, such as automatic mail sorting, statistical file digitalizing, bank notes recognition, etc. And the research progression in the research of handwritten numeral recognition can be promoted to many other fields, such as Chinese character recognition, human face recognition, etc. In a word, the research of handwritten numeral recognition has both of practical value and academic value.There are two key parts in OCR, feature extraction and divider design. In feature extraction part, Principal Component Analysis (PCA) is a generally adopted method. However, there are many problems within it, such as long computation period, low recognition rate and hard to be used in practical situation. Among the training algorithm of artificial neural network, BP algorithm is a widely used method. As BP algorithm is based on gradient descent theory, this algorithm has many problems such as local minimum, lack of theory for hidden layer neurons selection, long training period, etc.Recent years, a new algorithm, called 2D principal component Analysis is raised in human face recognition field. This method has higher recognition rate and shorter feature extraction time. This paper improved the feature generalization matrix based on 2DPCA by using new sample image pixel packaging strategy. This method improves system's recognition rate dramatically. This paper adopts a new neural network training algorithm, called algebra algorithm. And this method has a good performance in training and recognition section. The divider trained by this method has a better recognition rate, faster training time, more accurate than BP algorithm.1. Applied the 2DPCA algorithm to handwritten numeral recognition as feature extraction method. And the experimental result shows that 2DPCA has a better performance in feature extraction time and recognition rate than PCA.2. Raised a new feature matrix generalization method by using new sample image pixel packaging strategy, called NetPCA. This method combined two feature extraction methods, statistical and structural method. The best recognition rate base on this method is better than that of 2DPCA.3. Adopt a new neural network training algorithm, algebra algorithm, to train the neural network divider. This algorithm converts complex non-linear optimal problem to simple algebraic equations solving problem. And the cost function of the network that trained by this method can reach zero.4. Realize a handwritten number recognition system based on NetPCA and Algebra training algorithm. This system has passed the experiment on USPS database. The effectiveness and correctness of the algorithm are proved, too.