基于深度学习的CNN手写体数字识别

Handwritten Digit Recognition Based on the Deep Learning of CNN

从深度学习卷积神经网络(CNN)的结构功能出发,重点研究了手写体数字通过神经网络的识别性能,以及当手写体数字染有各种噪声(高斯噪声、椒盐噪声、泊松噪声和斑点噪声等)时的性能。为了改善和提高卷积神经网络的性能,采用迁移学习技术,将ReLu学习单元更换为LeakyReLu学习单元,其他层保持不变,其目的是为了进一步改善Sigmod神经元函数易饱和的缺点,提高了学习效率和速度。

Beginning with the structure and function of deep learning of CNN,the recognition properties of the handcraft script digits are chiefly investigated through CNN,and the recognition properties of the handcraft script digits coupled with noise(such as Gaussian,salt&pepper,poisson,speckle etc.).The technology of learning transfer is adopted to improve the properties of the designed CNN,replacing the learning unit ReLu with LeakyReLu,the other layers unchanged.It improves the defect that the sigmod function is easily saturated,and the learning efficiency.