基于局部保持的遥感场景零样本分类算法

Zero-Shot Classification for Remote Sensing Scenes Based on Locality Preservation

目标域遥感图像特征分布的变化,导致遥感场景零样本分类性能下降,针对该问题,提出一种基于局部保持的遥感场景零样本分类算法。首先,为减少冗余信息,采用解析字典学习方法,将源域中的场景图像特征和类别语义词向量嵌入到同一稀疏编码空间,并实现两者稀疏系数的强制对齐,以建立图像特征与词向量之间的关系;然后,通过保留图像特征空间中场景图像间的局部近邻关系,增强场景图像对应稀疏系数的鉴别性,以有助于对稀疏系数进行聚类分析;最后,为适应目标域图像特征分布变化,采用k-means算法对目标域场景图像的稀疏系数进行聚类,并以初始中心的类别标签作为对应的聚类簇中场景的类别标签。实验分别采用GoogLeNet和VGGNet图像特征,以数据集UCM作为源域遥感场景集,对目标域场景集RSSCN7进行零样本分类,获得了最高50.67%和53.29%的总体分类准确度,比现有算法各提升了8.06%和9.70%。实验结果表明:该算法能够适应目标域遥感场景图像特征分布的变化,显著提升遥感场景零样本分类效果,具有一定的优越性。

Due to the change of image feature distribution in target domain, the performance of zero-shot classification for remote sensing scenes degrades. To solve this problem, a zero-shot classification algorithm for remote sensing scenes based on locality preservation is proposed. Firstly, in order to reduce redundant information, the analysis dictionary learning method was exploited to embed the image features and word vectors of the source domain into the common sparse coefficient space, and the sparse coefficients were compulsively aligned for establishing the relationship between the image features and word vectors. Then, the discriminability of sparse coefficients of scene images was enhanced by preserving the local neighborhood relationship among scene images, which is helpful for clustering analysis on the sparse coefficients. Finally, in order to adapt to the change of image feature distribution, the k-means algorithm was utilized to cluster the sparse coefficients of scene images, and the class labels of the initial centers were used as the scene class labels. With the UCM remote sensing scene dataset as the source domain, zero-shot classification experiments were carried out on RSSCN7 scene dataset of the target domain via two type image features, i.e., GoogLeNet and VGGNet. The highest overall accuracies of 50.67% and 53.29% are obtained, which outperform the state-of-the-art algorithms by 8.06% and 9.70%, respectively. The experimental results show that this method can adapt to the feature distribution of remote sensing scenes, and significantly improve the zero-shot classification performance with certain advantages.