韩丽容¹, 李 鹏^1,2, 王廷伟¹, 李 慧¹, 任 鹏¹
1.中国石油大学(华东)信息与控制工程学院;2.数学工程与先进计算国家重点实验室

摘要:
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。
关键词:海洋观测图像快速检索卷积神经网络哈希算法FPGA
DOI：10.11759/hykx20171011024
分类号:
基金项目:数学工程与先进计算国家重点实验室开放基金(2017A05)和青岛市自主创新计划应用基础研究项目(16-5-1-11-jch)

Deep Hashing Retrieval of ocean observation images and its FPGA in-orbit implementation
HAN Li-rong,LI Peng,WANG Ting-wei,LI Hui,REN Peng
Abstract:
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.
Key words:Ocean imagesFast retrievalCNNHash algorithmFPGA





PDF全文下载地址:

点我下载PDF