基于深度度量学习的卫星云图检索

Satellite cloud image retrieval based on deep metric learning

针对传统云图检索方法难于获得理想的检索精度且检索效率低的问题,提出了一种基于深度度量学习的云图检索方法。首先设计了残差3D-2D卷积神经网络,以提取云图的空间及光谱特征。鉴于传统基于分类的深度网络所提取的特征可能存在类内差异大、类间差异小的问题,采用三元组训练网络,依据云图之间的相似性将云图映射到度量空间中,以使同类云图在嵌入空间中的距离小于非同类云图。在模型训练时,通过对无损三元组损失函数增加正样本对间距离的约束,改善了传统三元组损失的收敛性能,提高了云图检索的精度。在此基础上,通过哈希学习,将度量空间中的云图特征变换成哈希码,在保证检索精度的条件下提高了检索效率。实验结果表明,在东南沿海云图数据集和北半球区域云图数据集上,本文算法的平均精度均值(mean average precision,mAP)分别达到75.14%和80.14%,优于其他对比方法。

Due to the traditional cloud image retrieval methods are difficult to obtain ideal retrieval accuracy and retrieval efficiency,a cloud image retrieval method based on deep metric learning is proposed.Firstly,a residual 3D-2D convolutional neural network is designed to extract spatial and spectral features of cloud images.Since the features extracted by the traditional classify-based deep network may have greater differences intra-classes than inter-classes,the triplet strategy is used to train the network,and the cloud images are mapped into the metric space according to the similarity between cloud images,so that the distance of similar cloud images in the embedded space is smaller than that of non-similar cloud images.In model training,the convergence performance of traditional triplet loss is improved and the precision of cloud image retrieval is increased by adding a constraint on the distance between positive sample pairs to the lossless triplet loss function.Finally,through hash learning,the cloud features in the metric space are transformed into hash codes,so as to ensure the retrieval accuracy and improve the retrieval efficiency.Experimental results show that the mean average precision(mAP)of the proposed algorithm is 75.14%and 80.14%for the southeast coastal cloud image dataset and the northern hemisphere cloud image dataset respectively,which is superior to other comparison methods.