Handwritten Digit Recognition Based On ARM Platform And TensorFlow

Handwritten digit recognition technology is an indispensable and important part under pattern recognition,and its application prospects are very broad,which is of great significance for its research and development.The technology can be applied to hand-written digital automatic entry systems such as postal automatic sorting,financial statements,test score statistics,and financial digital statistics.Under the background of the comprehensive development and automation of global information technology,the demand for handwritten digit recognition technology has become very urgent.It is of great significance and value to study accurate and efficient recognition methods.With the continuous development of network information technology this year,deep learning has gradually become familiar and utilized by all,and the convolutional neural network has made a breakthrough in the field of computer science.Handwritten digit recognition systems are often designed on traditional PCs,which brings many problems such as high cost and inconvenient portability.If the handwritten digital system is transplanted to the ARM platform,these problems will be well solved.This paper mainly studies the construction of handwritten digital models by independent cyclic neural network and improved convolutional neural network and the design of handwritten digit recognition system on ARM platform.The main work is as follows:First,the traditional Recurrent Neural Network(RNN)has the problem of gradient disappearance and gradient explosion because of the parameter sharing in time.In this paper,a multi-layer Independent Recurrent Neural Network(IndRNN)is used to construct a handwritten digit recognition model,which can solve the gradient disappearance and gradient explosion problem of the traditional Recurrent Neural Network(RNN)by using the activation function such as Relu.The algorithm used in this paper is in MNIST.A good recognition rate was achieved on the data set.Secondly,this paper improves the convolutional neural network LeNet-5 model to construct a handwritten digital model.In order to improve the generalization ability and recognition rate of the network and avoid over-fitting,a Dropout layer is added after each convolution pooling layer to construct the network structure;in order to solve the non-convergence and the speed of convergence of the LeNet-5 model training recognition,using the Relu activation function instead of the usual Sigmoid activation function.Finally,in view of the high cost and inconvenient carrying of traditional PCs,this paper designs the handwritten digit recognition system on the ARM platform.The handwritten digits are taken by the camera provided by the ARM platform,the handwritten digits of the shooting set are image preprocessed,and the trained handwritten digit recognition model is transplanted onto the ARM platform,and finally the recognition result is obtained through model recognition.