摘要： 卷积神经网络(CNN)模型在图像识别中取得了良好的效果,但其识别精度还有进一步提升的空间。为此,设计一种新的图像识别模型CNN-GRNN。利用CNN提取样本图像中的多层次特征信息,将广义回归神经网络代替反向传播神经网络,以提高分类器的泛化能力和鲁棒性,通过均方差和降梯度法训练模型。基于COIL-100和手势库的实验结果表明,与灰度共生矩阵、HU距方法、CNN和CNN-SVM模型相比,CNN-GRNN的识别率分别提升了42.2%,13.43%,3.99%和1.86%,并具有较好的实时性。

关键词: 卷积神经网络, 广义回归神经网络, 支持向量机, 反向传播神经网络, 降梯度法

Abstract: Convolutional Neural Network(CNN) model has been widely used in image recognition.However,there is always room for further improvement of recognition accuracy.Aiming at this problem,a new model is proposed in this paper named CNN-GRNN.It utilizes CNN to extract hierarchical features from sample spaces,and then employs General Regression Neural Network(GRNN) to replace Back Propagation(BP) neural network to enhance generalization and robustness.In the end,it uses Root Mean Square(RMS) and gradient descent method to train the proposed model.Based on COIL-100 and gesture database,experimental results show that the proposed model is respectively improved by 42.2%,13.43%,3.99% and 1.86% in recognition rate compared with the Gray Level Co-occurrence Matrix(GLCM),HU invariant moments,CNN and CNN-SVM model.Therefore,it meets the real-time needs.

Key words: Convolutional Neural Network(CNN), General Regression Neural Network(GRNN), Support Vector Machine(SVM), Back Propagation(BP) neural network, gradient descent method

中图分类号: 

• TP391