Handwritten number recognition based on PCA and neural network

Abstract

In recent years, Artificial Neural Network (ANN) has been widely used in digital handwriting recognition by virtue of its strong fault-tolerant ability and classification ability. However, in traditional recognition methods, taking the number of image pixels as the input number of neurons will cause problems such as long learning and training time and low efficiency. This paper combines principal component analysis (PCA) algorithm with BP algorithm for handwritten digit recognition. After image preprocessing, PCA algorithm is used to reduce dimension of original data. The first 10 principal components, the first 30 principal components, the first 45 principal components, and the first 60 principal components are sent to the neural network as input neurons for training. Then, the test data set was used for testing. Finally, the simulation was analyzed from three dimensions: training efficiency, learning time and identification accuracy. The results show that when the number of input neurons is 60 and the number of hidden layer neurons is 30, the highest recognition rate is only 0.13% lower than that of the number of input neurons is 784, but the training time of the neural network is reduced by 94 seconds, and the efficiency is improved by 32%. When the number of hidden layer neurons is 50, the highest recognition rate is only 0.25% lower than that of input neurons is 784, but the time is reduced by 237 seconds and the efficiency is improved by 63%. This design solves the problem of low learning efficiency in digital recognition well.