基于Gabor核的卷积神经网络改进算法及应用

An improved convolutional neural network algorithm based on Gabor kernel and its application

针对Le Net-5网络模型识别分类准确度问题,提出一种基于深度Gabor卷积神经网络的识别分类方法。在Le Net-5模型的基础上,引入了Gabor层,使用Gabor核作为提取图像特征的卷积核,Gabor卷积核从图像频域的不同尺度、不同方向上提取更多特征。为了避免网络训练中的梯度消失问题,使用Relu函数作为网络中的激活函数。将改进后的模型在MNIST手写体数据集上进行试验,识别正确率达到99. 34%。与支持向量机和卷积神经网络等分类方法作比较,结果表明,改进后的深度Gabor卷积神经网络的具有更高的识别性能。

Aiming at the problem of classification accuracy of LeNet-5 network model,a recognition classification method is proposed based on deep Gabor convolution neural network.On the basis of LeNet-5 model,Gabor layer is introduced,and Gabor kernel is used as the convolution kernel to extract image features,which can extract more features from different scales and directions in image frequency domain.To avoid the problem of gradient disappearance in network training,the Relu function is used as an activation function in the network.The improved model was tested on the MNIST handwritten dataset,and the recognition accuracy was 99.34%.Comparisons with classification methods such as SVM and convolutional neural networks show that the improved Gabor convolution neural network with improved depth has higher recognition performance.