基于深度学习和复杂空间关系特征的多尺度遥感图像检索

Multi-scale remote sensing image retrieval based on deep learning and complex spatial relation

深度学习技术逐渐成为解决图像检索和图像分类问题的主流技术,然而现有算法不能有效地处理遥感图像中的复杂空间关系以及多尺度特征问题.为有效解决遥感图像检索问题,提出了一种综合考虑空间关系与尺度特征的新方法.首先,用深度学习方法检测过的遥感图像抽象为点集;其次,构造Delaunay三角网以描述全局空间关系特征;最后,在空间特征相同的部分中使用模糊性状模型匹配局部空间特征.该模型有效地规避了因尺度不同而产生的视觉差异.在UC Merced Land-Use和RS19等公开数据集上进行实验,结果表明,该算法在多尺度遥感图像检索精度方面的表现优于其他相关方法.

In recent years,the deep-learning technology has gradually become the solution to image retrieval and the mainstream technology of image classification.However,the existing algorithms cannot effectively deal with the complex spatial relations and multi-scale feature problems in remote sensing images.In order to solve the problem of remote sensing image retrieval effectively,a new method of considering spatial relation and scale feature is proposed.First,the remote sensing image detected by the depth learning method is abstracted as a point set.Secondly,the Delaunay triangulation is constructed to describe the global spatial relations feature.Finally,the fuzzy trait model is used to match the local spatial features in the same part of the spatial feature.The model effectively avoids the visual difference due to the different scale.Experimental results tested on some datasets such as UC Merced Land-Use and RS19 show that the algorithm is superior to other related methods in multi-scale remote sensing image retrieval accuracy.