摘要： 随着深度学习的不断发展，卷积神经网络（CNN）在目标检测与图像分类中受到研究者的广泛关注。CNN从LeNet-5网络发展到深度残差网络，其层数不断增加。基于神经网络中"深度"的含义，在确保感受野相同的前提下，给定标准的输入图片和输出特征图，对不同层数的卷积神经网络进行训练，并将训练结果与标准输出图进行对比。在此基础上，对标准的3×3卷积核进行分解，构建由2×2大小卷积核组成的CNN。根据目标特征是否具有中心对称的性质，提出多层卷积网络初始权值的选取规则。

关键词: 卷积神经网络, 卷积核, 感受野, 网络深度, 中心极限定理

Abstract: With the continuous development of deep learning,Convolutional Neural Network(CNN) have received extensive attention from researchers in target detection and image classification.CNN have evolved from LeNet-5 networks to deep residual networks,and the number of layers are increasing.According to the definition of the depth of a neural network,this paper trains CNN of different layers with given standard input images and an output feature graph.The training is performed on the premise that the CNN shares the same receptive field,and training results are compared with standard output graphs.On this basis,the standard 3×3 convolution kernel is decomposed,and a convolutional neural network composed of 2×2 convolution kernels is constructed.According to the central symmetry property of the target feature,the selection rule of the initial weight of the multi-layer convolutional network is proposed.

Key words: Convolutional Neural Network(CNN), convolution kernel, receptive field, network depth, central limit theorem

中图分类号: 

• TP391