基于特征深度融合的高分辨率遥感影像的检索

Image retrieval is an effective approach for object recognition in high resolution remote sensing images. However, there are some limitations in the traditional algorithms about feature extraction and feature fusion, which reduce the accuracy of image retrieval. First, image features extracted by the traditional methods have limited abilities to describe the characteristics of the object-oriented high-resolution remote sensing images. Convolutional neural networks (CNN) can be adopted for the high-level feature extraction. But it usually needs lots of labeled images for tuning the parameters of the networks. While the labeled images are limited, the accuracies of extracted features are degraded. Second, aiming to obtain rich information of remote sensing images, we need to take different types of image feature into consideration for image retrieval. However, current feature fusion methods usually lack of the unified feature space, which cannot fully elaborate the superiority of different types of features in image retrieval. In our project, for feature extraction, novel feature extraction method is developed with self-supervision and self-control. Its aim is to alleviate the dependence of image labels and enhance the performance of CNN. For feature fusion, with deep neural network and novel metric learning, different types of features are fused by means of unified feature space. The research can help us to extract the high-level features from the mass remote sensing images, and provide a high-precision and efficient solution to image retrieval. It is important to develop such a technical system for the development of image retrieval in the field of remote sensing.

图像检索是高分辨率遥感影像中地物识别的一种有效方法。传统特征提取、特征融合方法存在着局限性，降低了图像检索的精度。首先，传统方法提取的地物特征不能很好地表征面向对象的高分辨率遥感影像的特点，而卷积神经网络等高层次特征提取方法，模型提取特征的好坏依赖于给定图像标签的质量；第二，为了获取丰富的地物信息，需要考虑不同种类的地物特征，而目前不同特征的融合方法缺乏对统一的深度特征空间的定义，因而不能充分发挥不同图像特征优势。本项目为了强化CNN提取特征的能力，减弱CNN特征提取对图像标签的依赖，从特征的自我监督、自我判断角度入手，设计提取地物特征的算法，进行多尺度的高层次特征学习；通过引入新的度量学习模型，定义统一的特征空间，结合深度网络对多尺度特征进行深度融合，进而进行地物影像检索。研究成果对于发展完善遥感图像地物图像检索的技术方法体系，具有重要的科学意义与应用价值。

图像检索是高分辨率遥感影像中地物识别的一种有效方法。传统特征提取、特征融合方法存在着局限性，降低了图像检索的精度。本项目为了强化CNN提取特征的能力，减弱CNN特征提取对图像标签的依赖，从特征的自我监督、自我判断角度入手，设计提取地物特征的算法，进行多尺度的高层次特征学习；通过引入新的度量学习模型，定义统一的特征空间，结合深度网络对多尺度特征进行深度融合，进而进行地物影像检索。项目提出了一套用于图像检索的高层次特征学习和特征深度融合的关键算法，建立了多标签高分辨率遥感影像数据集，建立了高层次特征生成模型，并对生成的特征进行了检验，通过新的度量（关系）学习模型，定义统一的深度特征空间。在本项目的资助下，在IEEE Transactions on Geoscience and Remote Sensing、ISPRS Journal of Photogrammetry and Remote Sensing、IEEE Transactions on Multimedia等高水平期刊上发表论文19篇。本项目执行期间培养高级职称1名，博士生3名（已毕业1名），硕士生6人（已毕业2名）。研究成果对于发展完善遥感图像地物图像检索的技术方法体系，具有重要的科学意义与应用价值。