密集特征哈希的遥感场景分类

Remote sensing image classification based on DenseNet feature hashing

针对遥感场景的精准分类,提出了密集网络特征哈希的场景分类算法。基于密集网络输出的高层语义特征,经全连接层过渡降维后,激活函数生成归一化的特征向量作为分类层的判断输入,形成端到端的分类网络。训练后的网络作为特征提取器,将测试数据激活层特征映射生成二值哈希码,最后采用支持向量机分类。所提出的算法分别在UC Merced,WHU和NWPU-RESISC45公开数据集进行了验证,分别与传统局部特征描述子、迁移学习、深度特征编码3个层次的多种算法进行了对比,实验结果表明,相比于传统中低层语义特征,分类准确度得到大幅度提高;相比于深度学习网络的迁移,密集特征映射表达精细,聚集影像核心类别判断要素,更符合遥感影像的特征分布;相比于深度特征编码算法,特征结构简单,分类精度高,迁移和拓展性强,可以满足不同遥感场景分类要求。

To achieve accurate remote sensing scene classification,this study proposed a classification algorithm based on DenseNet feature hashing.First,dimension reduction was conducted for high-level semantic features output by a DenseNet through a fully connected layer.Then,normalized feature vectors were generated as the input of the classification layer using an activation function,and an end-to-end classification network was formed.Using the trained network as a feature extractor,the features of the activation layer of test data were mapped into binary hash codes.Finally,the remote sensing scene classification was conducted using support vector machine.The new algorithm was validated on public data sets UC Merced,WHU,and NWPU-RESISC45,and its classification effect was compared with that of multiple algorithms at three levels,namely the conventional local feature descriptor,transfer learning,and depth feature coding.The experimental results are as follows.The new algorithm had significantly higher classification accuracy than conventional algorithms based on mid-and low-level semantic features.Compared with the algorithm based on transfer learning,the proposed algorithm has fine-scale DenseNet feature mapping and accumulates elements used to determine core categories of images and,thus,is more suitable for the feature distribution of remote sensing images.Compared with the depth feature coding algorithm,the new algorithm has a simple feature structure,high classification accuracy,and strong transferability and extensibility and,thus,can meet the classification requirements of different remote sensing scenarios.