摘要
在标记样本较少的情况下,迁移学习对预训练模型的依赖度较高;当图像内容差异较大时,迁移学习的优势并不明显。因此,通过运用深度网络的非监督方法训练大量非标记样本,可以获得卷积神经网络的初始参数值。若卷积神经网络结构确定,网络的卷积层和全链接层需要确定参数,池化层和激活函数层并不需要设计参数,而全连接层属于特殊的卷积层,因此主要参数由卷积层确定。由于自编码可以复现输入数据的特征,因此卷积神经网络的每层卷积可使用自编码进行训练,从而得到本层的输入特征。文章对QuickBird的彩色合成图像进行了分类,其精度明显提高。
In the case of few labeled samples,transfer learning is highly dependent on the pre-trained model,when the image content is quite different,the advantage of transfer learning is not obvious.Therefore,the initial parameter values of the convolutional neural network can be obtained by training a large number of unlabeled samples using the unsupervised method of the deep network.If the structure of the convolutional neural network is determined,the parameters of the convolutional layer and the fully connected layer of the network need to be determined.The pooling layer and the activation function layer do not need to design parameters,and the fully connected layer is a special convolutional layer,so the main parameters are determined by the volume layering is confirmed.Since the self-encoding can reproduce the features of the input data,each layer of convolution in the convolutional neural network can be trained using the self-encoding to obtain the input features of this layer.The article classifies the color composite images of QuickBird,and its accuracy is significantly improved.
作者
陈蒙蒙
CHEN Mengmeng(Puyang Vocational and Technical College,Puyang,Henan 457100,China)
出处
《计算机应用文摘》
2022年第5期73-75,79,共4页
Chinese Journal of Computer Application
关键词
自编码预训练
卷积神经网络
迁移学习
self-coding pretraining
convolutional neural network
transfer learning
