海洋观测图像的深度哈希检索及FPGA在轨实现

Deep Hashing Retrieval of ocean observation images and its FPGA in-orbit implementation

近年来,星载/机载的在轨海洋观测设备日益增多。如何从星/机在轨观测的数据中快速且准确地检索出关键图像并进行回传,逐渐成为在轨实时处理的重要研究课题。哈希函数作为一种映射函数,能够将不同图像变换为相同长度的二进制码,并且保持类似图像二进制码的相似性。为了保证二进制编码的特征表达能力,采用深度卷积神经网络(Convolutional Neural Network,CNN)提取海洋图像的特征信息,对图像进行哈希变换。本文以海洋观测图像作为训练数据集,基于深度卷积神经网络在GPU(Graphics Processing Unit)上训练哈希算法。在此基础上,本文将训练好的哈希函数在FPGA(Field Programmable Gate Array)上实现,完成海洋观测图像的在轨快速检索,为海洋观测实时处理提供了有效技术手段。

Recently, the amount of in-orbit observation devices has greatly increased, and fast and accurate methods of retrieving ocean images observed by in-orbit satellites or UAVs have become more important. A Hashing algorithm, when used as a mapping function, represents different images with binary codes of the same length and renders similar images with the similar binary codes. To maintain the representational power of Hashing code, we employ a deep convolution neural network （CNN） to extract image features and conduct the Hashing transformation. Specifically, we use ocean images as training data and employ a deep CNN to train the Hash algorithm on Graphics Processing Unit （GPU）. We implement the trained Hash function on Field Programmable Gate Array （FPGA）, such that the ocean images can be quickly retrieved in orbit. The CNN hashing framework and its FPGA implementation scheme provide an effective means for processing marine information.