A Study On The Methods Of Handwritten Numeral Recongnition Based On Deep Learning

The research content of this paper is handwritten numeral recognition method based on deep learning.As the mainstream office mode carrier,computer makes the complicated and complicated office process electronic,digital and intelligent through information technology,which also makes people increasingly rely on computer.However,many of the work is still on paper due to the limitation of the on-site situation,which greatly increases the intensity of the work during the statistical analysis of the data,thus reducing the work efficiency.Therefore,we hope to combine the traditional way of writing office with artificial intelligence.In order to improve the problems of handwritten office,the handwritten numeral recognition method based on deep learning provides a solution.With the rapid development of artificial intelligence,the continuous research of convolutional neural network in the computer field makes deep learning increasingly known to people and has been widely applied in ordinary people's lives.However,due to the complexity and variability of handwritten digits,computer recognition has been challenged,and recognition with higher accuracy and accuracy is still the future development direction.This paper studies the efficiency of handwritten numeral recognition in deep learning under the premise of processing complex large samples.Currently deal with complicated large sample of data processing is not only dependent on professional high-performance equipment,the use of deep learning model complexity are also growing recognition accuracy can be ensured,and it increased the common usage scenarios of barriers to entry,thus reducing quantity and computing time,is a precondition for deep learning application more widely.In order to improve the identification efficiency,this paper proposes a scheme based on improving Inception structure and DenseNet structure,which can reduce the identification time of deep learning and improve the operation rate.The main contents of this paper are:1.Through theoretical analysis of the pros and cons of various deep learning frameworks,explore the improvement schemes based on Inception and DenseNet structures.2.Through experiments on various deep learning frameworks and analysis of experimental data,it is verified that after the convolution kernel size and parameter adjustment of the original Inception structure,the addition of DenseNet structure can reduce the operation time and improve the recognition efficiency without sacrificing the recognition accuracy.The experimental data showed that the training accuracy of the improved structure reached 99.47%,the verification accuracy reached 89.45%,and the training time of a single batch was 26.15%less than that of the model before the improvement.