基于深度卷积神经网络的遥感图像多层次特征表达及检索研究

Feature representation and information retrieval are challenging tasks in remote sensing field due to the characteristics of spatial-temporal complexity and massive diversity of remote sensing big data. Traditional remote sensing image retrieval methods are based on low-level features or combined low-level features. However, low-level features have the drawbacks of high dimension, weak discriminability and low retrieval accuracy. This project investigates multi-level feature representation and retrieval of remote sensing images based on deep convolutional neural networks, which mainly includes three parts: (1) We construct a multi-level feature representation model for remote sensing images based on the fact that remote sensing images can be regarded as the composition of pixels, regions, objects and scenes. In this multi-level feature representation model, images can be represented from pixels to feature patterns and visual objects; (2) Deep learning is introduced for automatic feature learning. The feature representation of each level is used as the input of deep convolutional networks to train the model; (3) The deep model can learn sparse features of remote sensing image via adaptive adjustment and persistent evolution, resulting in accurate feature representation and adaptive retrieval of remote sensing images. This project aims to make breakthroughs in the influence of feature grain on deep learning model performance, the fusion strategies of multi-level features as well as the collaborative optimization of multi-level feature and deep network, thereby providing a new way for deep learning based remote sensing image retrieval.

遥感大数据具有时空复杂性和海量多样性特点，其特征表达和检索是目前国际遥感科学技术的前沿，具挑战性。传统的基于低层特征的遥感图像检索方法，存在特征维数高、描述不完整、准确性差等缺点。本项目研究基于层次化卷积神经网络的遥感图像多层次特征表达及检索模型、理论和方法，内容包括：（1）针对遥感图像的像素—区域—对象—场景的层次结构，通过建立多层次的遥感图像特征描述，将遥感图像表示为从像素到特征模式和显著对象的层次模型；（2）将深度学习用于遥感图像检索，将不同层次的遥感图像特征描述作为深度卷积神经网络的输入进行模型训练；（3）通过自适应调整和不断演化学习遥感图像的稀疏特征，实现复杂遥感图像场景的准确特征描述和自适应检索。本项目通过特征粒度对深度学习网络性能的影响机理、不同层级特征的融合模型以及层次化特征与深度学习网络的协同优化三个关键科学问题上的突破，为基于深度学习的遥感图像检索提供一条新的解决途径。

遥感大数据具有时空复杂性和海量多样性特点，其特征表达和检索是目前国际遥感科学技.术的前沿，具挑战性。传统的基于低层特征的遥感图像检索方法，存在特征维数高、描述不完整、准确性差等缺点。本项目研究基于层次化卷积神经网络的遥感图像多层次特征表达及检索模型、理论和方法，内容包括：（1）针对遥感图像的像素—区域—对象—场景的层次结构，通过建立多层次的遥感图像特征描述，将遥感图像表示为从像素到特征模式和显著对象的层次模型；（2）将深度学习用于遥感图像检索，将不同层次的遥感图像特征描述作为深度卷积神经网络的输入进行模型训练；（3）通过自适应调整和不断演化学习遥感图像的稀疏特征，实现复杂遥感图像场景的准确特征描述和自适应检索。本项目通过特征粒度对深度学习网络性能的影响机理、不同层级特征的融合模型以及层次化特征与深度学习网络的协同优化三个关键科学问题上的突破，为基于深度学习的遥感图像检索提供一条新的解决途径。