基于Keras手写数字识别模型的改进

针对结构设计不合理的卷积神经网络导致MNIST识别的准确率低、收敛速度慢和训练参数多等问题,提出卷积神经网络结构的改进模型.改进的模型采用2次卷积、2次池化和3次全连接、采用Relu激活函数和Softmax回顾函数相结合,加入Dropout层防止过拟合,加入Flatten层优化结构.为了缩减代码量,采用API功能强大的Keras模型替代Tensorflow.对MNIST的训练集和测试集数据的准确率进行仿真实验,实验结果表明:采用改进的结构在MNIST的训练中不仅收敛速度快、训练参数少、损失率低,而且在测试集上的准确率达到99.54％、高于改进前的99.25％,对后续手写数字的研究具有重要意义.

Improvement of Handwritten Numeral Recognition Model Based on Keras

Aiming at the problems of low accuracy, slow convergence speed and many training parameters caused by the unreasonable structure design of convolutional neural network, an improved model of convolutional neural network structure is proposed. The improved model adopts 2-fold convolution, 2-pooling and 3-time full connection, combines relu activation function and softmax review function, adds dropout layer to prevent over fitting, and adds flatten layer to optimize the structure. In order to reduce the amount of code, tensorflow is replaced by keras model with powerful API. The experimental results show that the structure of this paper not only has fast convergence speed, less training parameters and low loss rate, but also has an accuracy rate of 99.54% on the test set, which is higher than 99.25% before the improvement. It is of great significance for the subsequent research of handwritten digits.