交叉点的神经网络识别及联机手写字符的概率神经网络识别初探

【作者】 郭力宾；

【导师】 吴微；

【作者基本信息】 大连理工大学 ， 计算数学， 2003， 硕士

【摘要】 人工神经网络发展到今天，已有五十多年的历史。在一代又一代学者的不懈努力下，不但理论基础逐渐充实、成熟，而且在信号处理、计算机视觉、模式识别、专家系统、工业控制与气象预测等许多领域有了广泛的应用。神经网络模式识别方法是近几年兴起的模式识别领域的一个新的研究方向。神经网络模式识别系统的研究，无论对神经网络理论的发展，还是对模式识别技术的实际应用，都有特别重要的意义。 BP网络是当前应用最广泛的一种人工神经网络。已被人们广泛的应用于神经网络模式识别(特别是图像及字符的识别)问题。 Kohonen提出的自组织特征映射(Self-Organizing Feature Map)神经网络(SOFM)因具有较强的拓扑组织能力和抗干扰能力，也广泛于应用于神经网络模式识别领域。 在1960年末被提出的HMM模型，已经被应用的连续的和演讲者无关的自动演讲识别中。近几年它也开始被广泛的应用于字符识别的工作中。 本文的工作主要是以下两个方面，都是侧重于有关神经网络的实际应用方面。 1．交叉点的神经网络识别。提出了一种识别灰度图像中交叉点的神经网络方法。用自组织特征映射网络和BP网络组成多分类器，识别图像中的二交叉点、三交叉点和四交叉点。在加入8％的干扰时，仍能达到85％左右的识别率。 2．概率神经网络用于联机手写字符的识别。由于计算机技术的高速发展，人们可以快速处理大量的数据，但数据的输入速度远远低于数据处理的速度，这大大妨碍了人们对计算机的使用。而且用键盘作为输入会打断正在进行的思维。所以研究一种方便的手写输入方法是很有必要的。到现在，单字符手写体识别，尤其是数字识别率接近95％，但对于连续的字符书写，尤其是有数学符号的情况下，由于联机手写的随意性比较大，写作不规范，还没有一个成熟的产品。本文应用HMM概率模型和神经网络结合，对联机手写数字和数学符号进行识别。对概率神经网络模型应用于联机手写识别进行了有益的探索。 与其他学者的一些工作相比，本文侧重于对使用概率神经网络的方法对联机手写进行识别，既利用了神经网络优异的分类性能，又避免了一般的神经网络实时性差的弱点。由于特征提取技术的局限性及算法的不成熟，本文的工作距离实际应用阶段还有很长的一段，有待进一步的改进完成。但是我们的试验结果表明，概率神经网络可以成功的用于联机手写字符的识别问题。更多还原

【Abstract】 The neural network has a history of over fifty years now. Generations of researchers have been making great efforts to build up its theoretical foundation and to apply it in many areas such as signal processing, machine vision, pattern recognition, expert system, industry control and weather forecast. In recent years, the pattern recognition based on neural networks has become a new active field. The study of the neural networks-based pattern recognition system is very important, not only to the development of neural networks theories, but also to the application of the pattern recognition techniques.The multi-layer feedforward back-propagation neural network has found a widely expanding range of applications. It has been used extensively for image and character recognition.The SOFM proposed by Kohonen is also used in the area of pattern recognition for its strong organizing ability on topology and its robustness.Since HMM was introduced at the end of 1960, it has been applied to the connected, speaker-independent, automatic speech recognition with the advantage of modeling various patterns. Recently it has also been widely adapted to character recognition.The work we have done is mainly focused on the following two problems.The first one is junction recognition using neural networks. Combining SOFM network and BP network we construct a multi-classifier to recognize junctions in images. When noises up to 8% are added, the system can still achieve a high recognition ratio of around 85%.Secondly, we use the HMM neural network to recognize on-line handwritten character. With the developing of computer technique, people can process data quickly. But the speed of typing data into computer is lower than the speed of processing data. It’s precluding people from using the computer. Furthermore, input with the keyboard would break thinking. So it’s necessary to find a convenient way of inputting. By this time, the rate of single handwritten character recognition, especially figure recognition, has reached 95%. But to the sequential characters, or mathematic character recognition there is not a perfect product. We recognize on-line handwritten figure and some mathematic characters using the HMM neural networks, and achieve a better result.Compared with other people’s work, we focus on recognizing characters with combination of HMM and neural networks. It makes use of the good performance of neural networks, and avoid the limitation that neural networks are not good at performing the real-time work. Due to the limitation of the characteristic selected and the algorithms used, the model we proposed still needs further improvement for practical application. But the results have shown that the HMM neural network can work successfully on the recognition of handwritten characters.更多还原