基于实例感知深度哈希学习的高分辨SAR图像检索

As an important remote sensing data, high-resolution Synthetic Aperture Radar (SAR) image plays an indispensable role in many areas, including environmental monitoring, national defense security, etc. With the development of imaging technology, the number of high-resolution SAR images increases dramatically. How to rapidly and accurately retrieve the proper content from a large amount number of images according to the user’s demands draws increasingly attentions. This project aims at developing the effective retrieval method for the high-resolution SAR image, in which the deep learning, hashing, and other outstanding techniques are integrated. First, a multi-task oriented deep feature-learning network is developed according to the properties of the high-resolution SAR image. The constructed network can suppress the speckle noise within the SAR image. In addition, the extract feature is strong enough to ensure the retrieval precision. Second, the hashing layer is embedded in the deep network to learn the hash code with the help of semantic similarities between SAR images. The obtained hash code is compact and robust, which can guarantee the retrieval efficiency. Finally, the verification system is constructed to test the effectiveness of our method using the different types of high-resolution SAR images. It is expected that the proposed method can provide the new idea for high-resolution SAR image retrieval task.

高分辨合成孔径雷达（Synthetic Aperture Radar，SAR）图像作为一种重要的遥感数据，在环境监测、国防安全等领域发挥着不可或缺的作用。随着成像技术的发展，高分辨SAR图像的数量与日俱增，如何快速、有效地从海量的图像中检索出用户关心的内容成为研究热点。本课题针对高分辨SAR图像内容检索任务中的关键问题，借鉴深度学习、哈希编码等技术，探索基于实例感知深度哈希学习的高分辨SAR图像检索新理论与新方法。课题首先根据高分辨SAR图像的特点，提出面向多任务的深度特征学习网络，提取抑噪、表征力强的视觉特征，保证检索的精度；其次，在深度特征学习网络的基础上加入哈希编码隐层，结合图像间的语义相似度，获得高分辨SAR图像紧致、鲁棒的哈希特征，保证检索的效率；最后，建立原型验证系统，在多种高分辨SAR图像上开展性能测试。期望通过研究，为高分辨SAR图像内容检索提供新的技术途径。

合成孔径雷达（Synthetic Aperture Radar，SAR）图像内容解译是遥感领域的重要问题，它在很多实际应用扮演者不可或缺的角色。如何从SAR图像想提取有用的信息，一直是摆在研究人员面前的难题。随着遥感技术的发展，SAR图像的数量激增，同时其分辨率也有大大幅度的提升，这为SAR图像解译工作带来了诸多挑战，如内容复杂、目标奇异性高等。本项目针对SAR图像内容挖掘、管理需求，借助基于内容的图像检索技术，开展了一系列科学研究，包括：特征提取、多示例分析、哈希编码等。相关研究内容在国际顶级期刊/会议上发表论文40余篇，授权/申请专利10余项。同时，研究成果成功应用于多项军事、民事项目中，为诸多行业（农业、环境等）提供了有效的技术支撑。