Applying Radam Method to improve treatment of Convolutional Neural Network on Banknote Identification

With the denomination recognition of existing convolutional neural network methods, a method which is based on a Radam optimizer and improved LeNet neural network has been proposed to further addressing problems in the accuracy and efficiency of banknote recognition. In order to improve the training efficiency of the system, the method of transfer learning has been used to extract and recognize the stability and accuracy of input data. For denomination pictures of different sizes, the image normalization method is adopted to realize imported standardized neural network data. In order to further improve the robust performance of the system algorithm, data enhancement methods are introduced, including the geometric transformation of the original data, adding noise, brightness, etc. At end the network connects the fully connected layer and the softmax layer to output labels with different denominations. Experiments show that the network meets the system's real-time requirements, and the algorithm has greatly improved the stability and accuracy of RMB denomination recognition with improvement from 95% to 99.97%. The experimental results illustrate that the algorithm structure is very practical.