%0 Journal Article %T 基于深度信念卷积神经网络的图像识别与分类<br>Image recognition and classification by deep belief-convolutional neural networks %A 刘琼 %A 李宗贤 %A 孙富春 %A 田永鸿 %A 曾炜 %J 清华大学学报(自然科学版) %D 2018 %R 10.16511/j.cnki.qhdxxb.2018.22.034 %X 针对基于卷积神经网络的图像识别采用随机初始化网络权值的方法易收敛到局部最优值的问题,该文提出了一种结合无监督和有监督学习的网络权值预训练算法。融合零成分分析白化与深度信念网络预学习得到的特征,对卷积神经网络权值进行初始化;通过卷积、池化等操作,对训练样本进行特征提取并使用全连接网络对特征进行分类;计算分类损失函数并优化网络参数。在公开图像数据库中进行了大量实验,与公开最佳算法比较,该算法在MNIST中的识别错误率降低了0.1%,在Caltech101中的分类准确率提升了0.56%,验证了该算法优于现有算法。<br>Abstract:Convolutional neural network (CNN) would easily converge to the local minimum if the network was randomly initialized in image classification tasks. A deep belief network pre-training method was developed by merging unsupervised and supervised methods. Feature sets were extracted from the image patches of zero component analysis (ZCA) whitening and deep belief pre-training to initialize weights of CNNs. Then, convolution features were extracted from the training samples by applying convolution and pooling operations and classified to a specific category through a fully connected network. Finally, the loss value was computed for global optimization. Extensive experimental evaluations on some public datasets show that this method is simple but very effective with the error rate decrease of 0.1% on MNIST and the accuracy increase of 0.56% on Caltech101, which indicates that this method is superior to similar methods. %K 深度信念网络 %K 图像识别 %K 卷积神经网络 %K < %K br> %K deep belief networks %K image recognition %K convolutional neural networks %U http://jst.tsinghuajournals.com/CN/Y2018/V58/I9/781