电子学报 ›› 2021, Vol. 49 ›› Issue (11): 2108-2116.DOI: 10.12263/DZXB.20200852

• 学术论文 • 上一篇    下一篇

基于新型联合感知矩阵的压缩学习目标识别技术

李峰1, 詹邦成1,2, 辛蕾1, 刘洋1, 刘志佳3, 肖化超4   

  1. 1.钱学森空间技术实验室,北京 100094
    2.河南大学河南省大数据分析与处理重点实验室,河南 开封 475004
    3.航天东方红卫星有限公司,北京 100094
    4.中国空间技术研究院西安分院,陕西 西安 710100
  • 收稿日期:2020-08-05 修回日期:2021-04-20 出版日期:2021-11-25 发布日期:2021-11-25
  • 作者简介:李 峰 男,1975年生,吉林人,钱学森空间技术实验室研究员,主要研究方向为遥感图像处理与重建、压缩感知、探测效能评估等. E‑mail:lifeng@qxslab.cn
    詹邦成(通信作者) 男,1995年生,河南信阳人. 河南大学计算机与信息工程学院,硕士研究生,主要研究方向为目标识别、图像处理. E‑mail:zbc2862333729@dingtalk.com
    辛 蕾 女,1988年生,辽宁阜新人,毕业于北京航空航天大学仪器科学与光电工程学院,硕士研究生,主要研究方向为压缩感知与空间信息智能提取. E‑mail:xinlei@qxslab.cn
    刘 洋 男,1992年生,山西忻州人,毕业于中国空间技术研究院西安分院,硕士研究生,主要研究方向为图像处理.E‑mail:liuandyang@aliyun.com
    刘志佳 男,1992年生,山西忻州人,毕业于中国空间技术研究院西安分院,硕士研究生,主要研究方向为图像处理. E‑mail:lzj303@sina.com
    肖化超 男,1985年生,湖南新邵人,毕业于中国空间技术研究院,工学博士,主要研究方向为空间数据处理.E‑mail:xiaohc@cast504.com

Target Recognition Technology Based on a New Joint Sensing Matrix for Compressed Learning

Feng LI1, Bang-cheng ZHAN1,2, Lei XIN1, Yang LIU1, Zhi-jia LIU3, Hua-chao XIAO4   

  1. 1.Qian Xuesen Laboratory of Space Technology,Beijing 100094,China
    2.Henan Key Laboratory of Big Data Analysis and Processing,Henan University,Kaifeng,Henan 475004,China
    3.DFH Satellite Co. ,Ltd. ,Beijing 100094,China
    4.China Academy of Space Technology (Xi’an),Xi’an,Shaanxi 710100,China
  • Received:2020-08-05 Revised:2021-04-20 Online:2021-11-25 Published:2021-11-25

摘要:

目标识别正逐渐成为自动化领域中提供准确目标类别信息的一项重要技术,并且当前大多数目标识别方法都是基于深度学习框架实现.通常,深度学习框架的输入数据均为原始图像数据,而在实际应用中,探测器获取原始图像数据并作为深度学习框架的输入进而实现目标识别的方式并非是高效的,数据获取并识别的过程包含了大量的冗余信息,降低了识别效率.在本文中,通过深度学习与压缩感知技术的结合,提出了一种基于联合感知矩阵的压缩学习目标识别技术(Target recognition technology based on a new joint sensing matrix for compressed learning,TRNPCL),使得探测器可快速生成目标图像多维压缩数据,且压缩数据可直接作为深度学习目标识别框架的输入数据,而无需再进行解压缩步骤.该方法不仅大大减小了深度学习框架的数据输入量,在与同等压缩比下的单空间域数据压缩学习方式相比较,还保持了较高的识别准确率.在未来,该方法有望成为一种更有效、更灵活的目标识别方法,并特别适用于指纹识别、人脸识别等应用领域.

关键词: 压缩感知, 深度学习, 目标识别, 压缩编码, 图像处理

Abstract:

Target recognition is gradually becoming an important technology to provide accurate target category information in the field of automation, and most current target recognition methods are based on machine learning frameworks. Generally, the input data of the machine learning framework is raw image data, but in practical applications, the way that the detector obtains the raw image data and uses it as the input of the deep learning framework to achieve target recognition is low efficient, contains a lot of redundant information, which reduces the recognition efficiency. In this paper, by combining machine learning and compressive sensing technology, a new target recognition method named as target recognition technology based on a new joint sensing matrix for compressed learning (TRNPCL) is proposed. Through the proposed method, the detector can quickly generate multi-dimensional compressed data of the target image, and the compressed data can be directly used as the input data of the deep learning target recognition framework without further decompression steps, which not only greatly reduces the amount of data input to the machine learning framework, but also maintains a higher recognition accuracy compared with the single-space domain data compression learning method under the same compression ratio. In the future, this method is expected to become a more effective and flexible target recognition method, and especially suitable for fingerprint recognition, face recognition and other application fields.

Key words: compressed sensing, deep learning, target recognition, compression coding, image processing

中图分类号: