投审稿平台
投稿指南
下载专区
地  址:北京市海淀区中关村科学院
南路6号中国科学院计算所342号 [地图]
《计算机辅助设计与图形学学报》编辑部
邮政编码:100190
电  话:010-62562491
     010-62600342
订阅信息
ISSN   1003-9775
CN    11-2925/TP
邮发代号:82-456
单  价:80.00元
全年订价:960.00元
订阅电话:010-64017032
在线期刊

空间调制全偏振计算成像场景迁移超分辨率方法

徐国明1,2,3), 袁宏武2,3)*, 薛模根3), 王峰3), 王杰1)
1) (安徽大学互联网学院 合肥 230039)2) (安徽新华学院大数据与人工智能学院 合肥 230088)3) (陆军炮兵防空兵学院偏振光成像探测技术安徽省重点实验室 合肥 230031)
分类号: TN911.73 DOI: 10.3724/SP.J.1089.2021.18699
出版年,卷(期):页码: 2021 , 33 ( 9 ): 1440-1449 徐国明
摘要: 空间调制型全偏振计算成像同时获得可见光和红外通道图像, 但是受探测器限制, 2个通道的图像空间分辨率不一致, 给后端的图像融合及目标探测过程带来不便. 为此, 提出基于场景特征迁移学习的空间调制型计算成像超分辨率方法. 首先在分析空间调制计算成像的相同场景异源图像特点基础上, 构建场景特征迁移模型; 然后建立改进场景迁移卷积神经网络结构并选择修正线性激活函数, 同时增加空间分辨率一致性约束; 再设计最优光谱迁移响应学习策略, 并作为前端输入加到超分辨率网络; 最后将光谱迁移响应优化与全偏振超分辨率重建的参数联合学习, 获得高分辨率偏振图像. 利用实际成像系统的仿真数据及系统数据进行2倍和3倍超分辨率实验, 从主观视觉效果、客观量化评价指标和偏振参量解析结果3个方面对多个指标进行评价. 结果表明, 文中方法在视觉效果上能够保持目标轮廓并抑制噪声干扰, 在16个客观指标对比数据上取得了10个优于、3个相同和3个低于的结果, 验证了该方法的有效性, 为成像系统定标校正提供数据支持.
关键词: 全偏振; 超分辨率; 计算成像; 迁移学习
Spatial Modulation Full Polarization Computing Imaging Super-Resolution via
Xu Guoming1,2,3), Yuan Hongwu2,3)*, Xue Mogen3), Wang Feng3), and Wang Jie1)
1) (School of Internet, Anhui University, Hefei 230039) 2) (School of Big Data and Artificial Intelligence, Anhui Xinhua University, Hefei 230088)3) (Anhui Province Key Laboratory of Polarized Imaging Detecting Technology, Army Artillery and Air Defense Forces Academy of PLA, Hefei 230031)
abstract: The spatial modulation full polarization computing imaging can obtain both visible and infrared channel images synchronously. However, due to the limitation of the detector, the spatial resolution of the two channels is inconsistent, which brings inconvenience to the following image fusion and target detection process. A spatial modulation computing imaging super-resolution method via scene feature transfer learning is proposed. Firstly, the scene feature transfer model is constructed. The model is based on the analysis of the spatial modulation computing imaging characteristics of different source images in the same scene. Secondly, the convolutional neural networks (CNN) structure of scene transfer is improved and the rectified linear unit activation function is selected. At the same time, the spatial resolution consistency constraint is added. Then, the optimal spectral transfer response learning strategy is designed and added to the super-resolution network as the front-end input. Finally, the parameters of spectral transfer response optimization and full polarization super-resolution reconstruction are learned together to obtain the high-resolution polarization image. Some super-resolution experiments with scale factors 2 and 3 are carried out using simulation data and system data of the actual imaging system. The experiments result is evaluated by multiple indicators with three aspects such as subjective visual effect, objective quantitative index and polarization parameter analysis results. In the visual effect, the method can keep the object contour and suppress the noise interference. On the 16 objective index data, the method obtains 10 better, 3 equal and 3 lower results compared with the others. The results verify the effectiveness of the method and also provide data support for calibration correction of imaging system.
keyword: full polarization; super resolution; computing imaging; transfer learning
 
Copyright © 2004《计算机辅助设计与图形学学报》版权所有
电话:010-62600342 传真:010-62562491
E_mail:jcad@ict.ac.cn