基于深度学习的编码孔径光谱成像重构算法  

Reconstruction algorithm of coded aperture spectral imaging based on deep learning

作　　者：蒋伊琳[1] 张建峰 吴进 JIANG Yi-lin;ZHANG Jian-feng;WU Jin(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China;School of Information Engineering,Liaoning Vocational College of Ecological Engineering,Shenyang 110122,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,哈尔滨150001 [2]辽宁生态工程职业学院信息工程学院,沈阳110122

出　　处：《吉林大学学报（工学版）》2020年第6期2221-2228,共8页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(61571149).

摘　　要：针对编码孔径光谱成像系统中利用传统算法重构高光谱图像时存在计算复杂度高、图像质量不好等问题,本文提出了一种基于深度卷积神经网络的高光谱图像重构方法。具体包括两个步骤:首先,对压缩测量值进行预处理得到目标图像的初始值,为此构建新的编码孔径模板;然后,在初始值与真实图像之间建立深度卷积神经网络,通过深度学习得到映射关系,再利用训练的模型重建光谱数据立方体。通过仿真实验,将本文方法与梯度投影稀疏重构(GPSR)、正交匹配追踪(OMP)以及两步迭代收缩阈值(TwIST)算法进行了对比,结果表明,利用本文方法重构高光谱图像耗费的时间大大减少,所得到的图像更加清晰,峰值信噪比和结构相似性均有明显提高。In the coded aperture spectral imaging system,the traditional algorithms have caused high computational complexity and poor reconstruction quality when reconstructing hyperspectral images. To solve these problems,a hyperspectral image reconstruction method based on deep convolutional neural network is proposed in this paper. The specific process consists of two steps. Firstly,the compressed measurements are preprocessed to obtain the initial value of the target image. For this purpose,a new coded aperture template is built. Then,a deep convolutional neural network between the initial value and the real image is established,and the mapping relationship is obtained through training. The trained model is applied to reconstruct the target spectral data cube. The experiment results are compared with the results of gradient projection sparse reconstruction(GPSR),orthogonal matching pursuit(OMP)and two-step iterative shrinkage/thresholding(TwIST)reconstruction algorithms. The time to reconstruct hyperspectral images using this new method is greatly reduced and the obtained image is clearer. Also,the peak signalto-noise ratio and structural similarity are significantly improved.

关 键 词：信息处理技术 编码孔径光谱成像 深度学习 卷积神经网络 高光谱图像重构 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]