Using convolutional features and a sparse autoencoder for land-use scene classification
Using convolutional features and a sparse autoencoder for land-use scene classification
复制标题
DOI：
10.1080/01431161.2016.1171928
复制
发表时间：
2016-05
影响因子：
3.4
通讯作者：
Esam Othman;Y. Bazi;N. Alajlan;H. Alhichri;F. Melgani
中科院分区：
文献类型：
--
作者：
Esam Othman;Y. Bazi;N. Alajlan;H. Alhichri;F. Melgani
摘要在本文中，我们提出了一种基于卷积特征和稀疏自动编码器（AE）的新方法，用于场景级别的土地使用（LU）分类。该方法首先从深度卷积神经网络（CNN）中生成正在从辅助域中进行大量标记数据的深度卷积神经网络（CNN）的初始特征表示。然后，将这些卷积特征作为稀疏AE的输入，用于以无监督的方式学习新的合适表示形式。在此预训练阶段之后，我们提出了两种不同的方案来构建分类系统。在第一种情况下，我们在AE编码层的顶部添加了一个SoftMax层，然后使用手头可用的目标训练图像以监督的方式微调所得网络。然后，我们根据SoftMax层提供的后验概率对测试图像进行分类。在第二种情况下，我们从重建角度查看分类问题。为此，我们训练多个类别的AE（即每个AE一个AE），然后根据重建误差对测试图像进行分类。与最先进的方法相比，在加利福尼亚大学（UC）梅塞德大学和Banja-Luka Lu公共数据集上进行的实验结果证实了该方法的优越性。
ABSTRACT In this article, we propose a novel approach based on convolutional features and sparse autoencoder (AE) for scene-level land-use (LU) classification. This approach starts by generating an initial feature representation of the scenes under analysis from a deep convolutional neural network (CNN) pre-learned on a large amount of labelled data from an auxiliary domain. Then these convolutional features are fed as input to a sparse AE for learning a new suitable representation in an unsupervised manner. After this pre-training phase, we propose two different scenarios for building the classification system. In the first scenario, we add a softmax layer on the top of the AE encoding layer and then fine-tune the resulting network in a supervised manner using the target training images available at hand. Then we classify the test images based on the posterior probabilities provided by the softmax layer. In the second scenario, we view the classification problem from a reconstruction perspective. To this end we train several class-specific AEs (i.e. one AE per class) and then classify the test images based on the reconstruction error. Experimental results conducted on the University of California (UC) Merced and Banja-Luka LU public data sets confirm the superiority of the proposed approach compared to state-of-the-art methods.