ニューラルネットワークを用いた高速搬送紙幣の織別

抄録

In conventional bill money recognition machines, we develop the recognition algorithm according to the transaction speed and diffierence of various specifications. However, development of the algorithm for the recognition has been based on the trial and error method. Many researchers have reported that neural networks are suitable for pattern recognition because of the ability of selforganization, parallel processing, and generalization. In this paper, we present a new bill money recognition method with neural network and show the effectiveness of the present algorithm compared with the conventional method by discrimination functions. Furthermore, we transform bill money data by FFT into frequency domain to reduce influence of the noise due to conveyed fluctuation. Then we adopt these Fourier coefficients or its amplitudes as input of the neural network. We show that the ability of recognition can be evaluated in detail by introducing a new measure of reliability.