计算机与现代化

• 图像处理 • 上一篇    下一篇

基于CNN回归网络的遥感图像颜色自动校正

  

  1. (1.中国科学院大学电子电气与通信工程学院,北京100190;2.中国科学院电子学研究所,北京100190; 3.中国科学院空间信息处理与应用系统技术重点实验室,北京100190;4.中国国际工程咨询公司,北京100048)
  • 收稿日期:2017-04-12 出版日期:2017-12-25 发布日期:2017-12-26
  • 作者简介:杜尚尚(1992-),女,山东邹城人,中国科学院大学电子电气与通信工程学院硕士研究生,研究方向:深度学习在遥感图像领域的应用; 雷斌(1978-),男,研究员,博士,研究方向:遥感地面应用系统; 郭嘉逸(1992-),男,博士,研究方向:SAR图像处理; 卢晓军(1977-),男,江苏泰州人,中国国际工程咨询公司高级工程师,博士后,博士,研究方向:智能控制,信号处理。
  • 基金资助:
    国家自然科学基金资助项目(61331017)

A Method of Automatic Color Correction for Remote Sensing Image  Based on CNN Regression Network

  1. (1. School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100190, China;  2. Institute of Electrics, Chinese Academy of Sciences, Beijing 100190, China; 3. Key Laboratory of Spatial Information Processing and Application System, Chinese Academy of Sciences, Beijing 100190, China; 4. China International Engineering Consulting Corporation, Beijing 100048, China)
  • Received:2017-04-12 Online:2017-12-25 Published:2017-12-26

摘要: 目前有很多成熟而有效的图像色彩校正算法,但是针对海量的遥感影像数据尚未有行之有效的大批量自动化色彩校正方法。为了解决这个问题,本文提出一种基于CNN(Convolutional Neural Networks)回归网络的遥感图像颜色自动校正方法——ACCN(Auto Color-Correction Network)模型。该模型通过预测消除遥感图像颜色误差后的RGB通道颜色分布直方图实现其色彩的自动校正。ACCN模型由20000幅GF-1号遥感图像在Tensorflow框架上以KL散度为损失函数训练获得。通过测试样本验证,经过该模型校正后,图像色调和谐,色彩鲜明。实验结果表明,该方法在校正过程中图像无需任何预处理,无需手动调节参数,无需参考图像,适宜遥感图像大规模自动化的颜色校正。

关键词: 遥感图像, 卷积神经网络, 颜色校正, 颜色直方图, 自动化

Abstract: At present, there are a lot of mature and effective image color correction algorithms. However there isn’t an effective automatic color correction method for massive remote sensing image data. In order to solve the problem, this paper proposes a fully automatic method named ACCN(Auto Color Correction Network)to correct multispectral remote sensing images’ color based on CNN (Convolutional Neural Network). The model corrects the multispectral remote sensing images’ color by predicting its’ Ground-truth color histogram of each color channel. The model is trained on 20 thousand pieces of remote sensing images in the Tensorflow framework. We get the optimal model which realizes automatic color correction for multispectral remote sensing image through many times repeated fine-tuning. Experiments show that the multispectral remote sensing image through automatic color correction becomes harmonious and flaming. This method can correct the color of large-scale remote sensing images automatically. 

Key words: remote sensing image, CNN, color correction, color histogram, automation

中图分类号: