基于词向量一致性融合的遥感场景零样本分类方法

Zero-Shot Classification Method for Remote-Sensing Scenes Based on Word Vector Consistent Fusion

遥感场景类别的语义词向量与图像特征原型的距离结构不一致问题,严重影响遥感场景零样本分类效果。针对该问题,利用不同词向量间一致性,提出一种基于解析字典学习的语义词向量融合方法,以提升遥感场景零样本分类效果。首先,采用解析字典学习方法,提取场景类别的不同词向量的公共稀疏系数,并作为融合后的语义词向量;然后,同样采用解析字典学习方法,将场景类别的图像特征原型嵌入到融合后的词向量空间,与融合后的词向量进行结构对齐,降低距离结构的不一致性;最后,通过联合优化获得未知类的图像特征空间类别原型表示,并采用最近邻分类器完成未知类别遥感场景的分类。在3种遥感场景数据集和多种语义词向量上进行定量和定性实验。实验结果表明,通过词向量融合可以获得与图像特征原型结构更一致的语义词向量,从而显著提升遥感场景零样本分类的准确度。

The problem of distance structure difference between the word vectors and visual prototypes of remote-sensing scene classification seriously influences the performance of the zero-shot scene classification.Herein,a fusion method based on analytical dictionary learning is proposed to exploit the consistency among the different kinds of word vectors for the performance improvement of the zero-shot scene classification.Firstly,the common sparse coefficients of different kinds of word vectors of scene classification are extracted by analytical dictionary learning method and acted as the fused word vector.Secondly,the visual prototypes are embedded into and structure-aligned with the fused word vector by analytical dictionary learning method similarly,to reduce the distance structure inconsistency.Finally,the prototypes of the unseen classes in the image feature space are obtained via joint optimization,and the nearest neighbor classifier is used to complete the classification of remote-sensing scenes from the unseen classes.Quantitative and qualitative experiments are also conducted on three remote-sensing scene datasets with the fusion of various word vectors.The experimental results show that the fused word vector is more structure-consistent with the prototypes in the image feature space,and the zero-shot classification accuracies of the remote-sensing scenes can be significantly improved.