基于稀疏卷积核的卷积神经网络研究及其应用

摘要

In view of the convolutional neural networks (CNNs) training image data,the learned convolution kernel is no rules,this paper proposes a convolutional neural network algorithm based on sparse convolution kernel.The method is adding sparse constraints to the squared-error loss function.When correcting convolutional kernel in back propagation,it can obtain the convolutional kernel that similar to the first order differential gradient operator,in where the part of numerical value is 0 or tends to 0,so it can be better to extract image edge features.According to the experiment of training the sign language image data and the license plate image data,the results show that the part of the convolution kernel has the form approximate to first order differential gradient operator and the convergence speed of the network learning is faster than the CNN.%针对卷积神经网络训练图像数据时,其学习到的卷积核是杂乱无章,没有规则的,提出了基于稀疏卷积核的卷积神经网络算法.该方法通过对平方误差代价函数加入稀疏约束项,在反向传播中修正卷积核时,使其学习到的部分卷积核近似于一阶微分梯度算子,即学习到的卷积核中部分值是0或者趋于0,可更好地来提取图像边缘特征.通过对手语图像数据及车牌图像数据进行训练的实验结果显示,其学习到的部分卷积核具有近似一阶微分的模板形式;并且相对经典卷积神经网络,该算法的识别正确率有所提高.