摘要
深度信念网络可以通过低层特征组合抽象形成更高层的特征,具有良好的学习能力,其无监督学习特点减少了人工劳动量。主要研究了使用深度信念网络(DBN)对手写数字进行识别的方法,实验表明设计识别数字的网络模型时的最佳层数为5层。使用MNIST数据库中的60 000个图片训练深度信念网络,再使用MNIST中的另外10 000个图片测试网络,得出高达93.42%的准确率,高于同等条件下的SVM。另外,在深度学习网络中引入Dropout参数,可以在使用少量样本的情况下获得更高的识别准确率。
By means of combining features of image low level, deep belief network(DBN) can form the feature of higher level, which has a good learning ability. Also, carrying out unsupervised learning can reduce the amount of manual labor. The use of deep belief networks to recognize handwritten numerals is mainly studied. When designing the network model, the optimal number of layer is 5. By using 60 000 images in the MNIST database to train the deep belief network, and then using the rest MNIST 10 000 pictures testing network, its recognition accuracy can reach 93.42%, which is higher than that of the SVM. Moreover, by using the Dropout training techniques in DBN, the network using a small amount of sample training has higher recognition accuracy than that of the network of Dropout absence.
出处
《工业技术创新》
2016年第5期921-924,共4页
Industrial Technology Innovation
