Classification of New and Used Bills by Acoustic Data Using Neural Networks

在这个项目中，我们提出了一种利用神经网络实现票据通过银行机器时，从声学数据中对新旧票据进行智能分类的方法。本课题是综合各类神经网络的智能分类器系统。特别地，我们采用了三种神经网络，即基于误差反向传播算法的分层网络、自组织映射网络和学习向量量化网络。为了完成本课题的研究，我们采用以下方法对这些智能分类系统进行了综合：(1)新旧票据声学数据的特征提取利用频谱和倒谱数据提取新旧票据声学数据的具体特征。然后利用自组织映射神经网络将这些数据分为新账单类别和旧账单类别。(2)基于遗传算法的网络规模优化为了增强网络的泛化能力，我们将遗传算法应用于分层神经网络和竞争学习网络，以获得最小的代价。然后我们可以在一些约束条件下将网络的大小减小到尽可能小的程度。(3)学习向量量化分类为了从频谱和倒谱上对新票据和旧票据进行分类，我们训练了基于学习向量量化的神经网络。然后对真实账单的测试数据集进行分类，得到90%以上的分类结果。(3)系统硬件实现为了加快算法的计算速度，降低成本，我们开发了系统硬件。从而实现廉价、高速的硬件系统，实现票据分类的规范。

In this project, we have proposed an approach to realize an intelligent classification of new and used bill money from acoustic data via banking machines when bills are passed in those machines by using neural networks. The present project is to synthesize an intelligent classifier system based on various types of neural networks. Especially, we have adopted here three kinds of neural networks which are a layered network by the error back-propagation algorithm, a self-organizing map network, and learning vector quantization networks. To complete the project study, we have adopted the following approach to synthesize those intelligent classification systems :(1) Feature Extraction of Acoustic Data from New and Used BillUsing spectrum and cepstrum data, we have extracted the specific features of acoustic data obtained from the new and used bill money. Then using the self-organizing map neural network, we have classified those data into two classes which mean new bill category and used bill category.(2) Optimization of Network Size by Genetic AlgorithmsTo enhance the generalization of the networks, we have applied the genetic algorithms to the layered neural networks and competitive learning networks such that the minimum cost could be obtained. Then we could reduce the network size as small as possible under some constraints.(3) Classification by Learning Vector QuantizationTo classify the acoustic data from new and used bills from the spectrum and cepstrum, we have trained the neural networks based on the learning vector quantization. Then we could obtain more than 90% classification results for test data set from real bills.(3) Hardware Implementation of the Proposed SystemTo speed up the computation of the proposed algorithm and reduce the cost, we have developed the hardware of the proposed system. Then we could realized cheep and high speed hardware system to achieve the specification of the bills classification.