稠密特征编码的遥感场景分类算法

Remote Sensing Image Classification Based on Dense Feature Coding

针对遥感影像快速有效的场景分类,提出了一种低维度稠密特征编码的场景分类算法.首先提取遥感图像不同尺度下的稠密特征,利用Hellinger kernel对原始特征进行映射变换形成新的特征空间,采用主成分分析对新的特征降维并进行Fisher编码量化,进而实现遥感图像的低维度稠密特征表达,最后在线性支持向量机中完成遥感影像的场景分类.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.实验结果表明,作为一种改进的中层语义特征表达算法,相比于传统中低层语义特征,分类准确度得到大幅度提高,相比于深度学习算法,所提算法能够有效兼顾计算复杂度和分类准确率,实现不同指标间良好的平衡,满足遥感场景分类的实用性要求.

To classify the scenes in remote sensing images efficiently and effectively,we propose a low-dimensional dense feature coding based scene classification method.Firstly,we extract dense features from remote sensing images at multiple scales and map the original features to a new feature space by using Hellinger kernel.Then,we perform dimensionality reduction and quantization for the obtained features by using PCA and Fisher coding to obtain low-dimensional dense features of the sensing image.Finally,we classify the scenes by using a linear support vector machine.We validate the performance of the proposed algorithm by conducting extensive experiments on three public scene classification datasets.Experimental results demonstrate the following advantages 1) compared with the traditional low or middle-level semantic feature,the classification accuracy has been significantly improved;2) compare with the deep learning-based method,the proposed method can make a good trade-off between computational complexity and classification accuracy,and meet the practical requirements.