The Implementation Of A System Of Handwritten Numeral Recognition With CNN Based On FPGA

Handwritten numeral recognition is an important branch of optical character recognition technology.How to use electronic devices to automatically identify the handwritten Arabic number is the research object.Due to the small number of categories for handwritten numeral recognition,it is helpful to make an in-depth analysis and verify some new theories.The artificial neural network(ANN)is the most obvious example.A considerable part of the ANN model and algorithms are used as a specific experimental platform for handwriting recognition.Convolution Neural network(CNN)is a kind of artificial Neural network,is a variant of multilayer perceptron model,evolved from the concept of biology,has become a current research hot spot in the field of speech analysis and image recognition.This paper completed the convolutional neural network applied to handwritten numeral recognition algorithm in Matlab.The convolutional neural network uses a 7 layer structure,respectively as follows: Input Layer,Convolution Layer,Pooling Layer,Convolution Layer,Pooling Layer,Full-Connection Layer,Output Layer.The convolutional neural network algorithm can be divided into four stages: The first stage is the initialization of convolutional neural network structure and parameters;The second stage is the training of the convolutional neural network parameters;The third stage is the test of convolutional neural network recognition accuracy;The fourth stage is the extraction and storage of the convolutional kernels and weighting parameters.Further we finished the implementation of the convolutional neural network based on FPGA.First we complete the design of the whole hardware structure of the convolutional neural network and the division of function modules.In implementation,the first step is to transform the data of matlab algorithm into fixed-point data and find the right precision.As a result of the limitation of resources on the FPGA chip,we need reuse the convolution module in the system.Finally the validation of the convolutional neural network system based on FPGA need to be done.The training and testing of convolutional neural network both use MNIST data set.There are 60000 training images and 10000 testing images in this data set.The Matlab algorithm and the system on FPGA both can achieve higher recognition accuracy.