Abstract: On the issues about the slow convergence speed and low identification rate in handwritten digit recognition, based on CNN（Convolutional Neural Network） in which the convolution kernels are initialized randomly, an improved algorithm in which the convolution kernels are initialized by PCA（Principal Component Analysis） is proposed. Firstly, training samples are selected and sent to CNN. In the corresponding layer feature map is processed by image block extraction, after that feature vectors extracted by the way of PCA in layered learning are used to initialize the convolution kernels. Finally, original images of the CNN are processed by these convolution kernels. Compared with the CNN handwritten digit recognition algorithm that randomly initializes the convolution kernel, the improved algorithm not only converges when applied to the MNIST database training, but also has fewer iterations and higher recognition rate when the same mean square error is generated.

Key words: Principal Component Analysis（PCA）, Convolutional Neural Network（CNN）, convolution kernel, handwritten digit recognition

摘要： 针对卷积神经网络对手写数字识别训练在卷积核随机初始化情况下收敛速度慢和识别率低的问题，提出一种主成分分析（PCA）初始化卷积核的卷积神经网络（CNN）手写数字识别算法。算法首先选取训练样本集并将其送入CNN，在相应层对Feature Map进行全覆盖取图像块处理，然后进行分层PCA学习，将学习到的特征向量做为对应卷积层的卷积核参数进行初始化，最后再用这些卷积核对原始图像进行卷积操作。实验结果表明，与随机初始化卷积核的CNN手写数字识别算法相比，改进的算法在应用MNIST数据库训练时不仅收敛，而且在产生相同均方误差的情况下迭代次数少，识别率高。

关键词: 主成分分析, 卷积神经网络, 卷积核, 手写数字识别