摘要
光谱图像含有丰富的空间和光谱信息,能够反映目标的组成、结构和材料特性,在航天遥感、医疗诊断和机器视觉等领域具有重要的应用价值.近年来,光谱成像技术作为热点研究领域受到广泛关注.传统光谱成像技术采用沿空间维度或光谱维度扫描的方式,依次获取待测物体表面的光谱信息.由于曝光时间较长,传统光谱成像技术不适用于拍摄动态场景.编码孔径快照光谱成像(Coded Aperture Snapshot Spectral Imaging,CASSI)是光谱成像的前沿技术方案,能够从单次曝光中快速获取动态场景的光谱图像,其包括两个阶段:对高维光谱图像的“编码降维采集”和对低维观测图像的“解码升维重构”.CASSI的早期研究聚焦于“编码降维采集”,通过物理系统设计提升图像编码的有效性,包括编码模板设计和双相机系统设计.目前,CASSI的“编码降维采集”物理系统趋于稳定,其“解码升维重构”决定了光谱成像的质量和效率.本文综述CASSI的重构算法.首先介绍CASSI的物理系统和前向模型,详细描述物理系统的组成元件和硬件参数,推导CASSI前向模型的数学表达;其次梳理CASSI重构的特点和挑战,其挑战主要存在于系统前向模型、先验表示模型、算法灵活性、算法复杂度、实物数据集等方面;之后重点归纳重构算法的研究现状,包括基于优化模型的重构算法和基于深度学习的重构算法.基于优化模型的重构算法利用凸优化模型求解线性逆问题,结合平滑、稀疏、低秩等手工设计的先验表示模型降低逆问题的欠定性;基于深度学习的重构算法利用数据驱动的方式建立先验表示模型,结合端到端全网络、深度展开、即插即用等框架求解重构图像.接着比较主流算法的重构质量和计算效率,以峰值信噪比、结构相似度、光谱角制图为重构质量的评价指标,以模型参数量、浮点计算量为计算效率的评价指标.最后讨论现有工作的不足和未来研究趋势,指出当前仍未解决的领域痛点,展望进一步的研究方向,为本领域开拓创新提供参考.
Spectral images contain a wealth of spatial and spectral information,enabling them to effectively reflect an object's composition,structure,and material properties.They have significant application value in aerospace remote sensing,medical diagnosis,and machine vision.In recent years,spectral imaging technology has got significant attention and emerged as a prominent research area,Conventional spectral imaging techniques scan along the spatial or spectral dimensions,enabling the sequential acquisition of spectral information from the object's surface.Due to the extended exposure time,these techniques are unsuitable for capturing dynamic scenes.Coded Aperture Snapshot Spectral Imaging(CASSI)is a cutting-edge technique for spectral imaging that allows for the rapid acquisition of spectral images of dynamic scenes from a single exposure.It consists of"encoding dimension-reduction collection"of high-dimensional spectral images and"decoding dimension-increase reconstruction"of low-dimensional measurements.Early research on CASSI primarily focused on the"encoding dimension-reduction collection"stage,aiming to enhance the effectiveness of image encoding through physical system design,including the design of coded aperture and dual-camera system.At present,the physical system for the"encoding dimension-reduction collection"stage has become fixed.The"decoding dimension-increase recon-struction"stage plays a crucial role in determining the quality and efficiency of spectral imaging.This paper presents a comprehensive overview of CASSI reconstruction algorithms,which aims to provide readers with a detailed understanding of the inner workings and intricacies of the various algorithms.First,we introduce the physical system and forward model of CASSI,providing a detailed description of the components and hardware parameters of the physical system and deriving the mathematical expression of the CASSI forward model.Second,we outline the characteristics and challenges of CASSI reconstruction,which mainly contain the forward model,prior represen-tation model,algorithm flexibility,algorithm complexity,and real-world datasets.Next,we summarize the current research status of reconstruction algorithms,including optimization-based and learning-based reconstruction algorithms.The optimization-based reconstruction algorithms employ convex optimization models to address the challenging linear inverse problems effectively.These algorithms apply crafted prior representation models,including but not limited to smoothness,sparsity,and low-rank,to tackle the inverse problem's inherent ill-posedness.The learning-based reconstruction algorithms take a different data-driven approach to establishing prior representation models.These algorithms leverage the power of deep learning frameworks,such as end-to-end networks,deep unfolding,and plug-and-play,to effectively solve the reconstruction problem.With the capabilities of deep learning,these algorithms can learn and fit the underlying patterns and structures within the data,leading to enhanced reconstruction performance.By contrasting the optimization-based and learning-based methods,we comprehensively understand the diverse methodologies of CASSI reconstruction to explore the inner workings and potential benefits and limitations.Furthermore,a thorough comparison is conducted utilizing various evaluation metrics to assess mainstream algorithms'reconstruction quality and computational efficiency.These metrics include peak signal-to-noise ratio,structural similarity,and spectral angle mapping for evaluating the reconstruction quality.In addition,model parameter count and floating-point operations are utilized to measure the computational efficiency.Finally,the shortcomings of existing work and future research trends are discussed.The unresolved pain points in the current field are identified,and potential research directions are highlighted,providing valuable insights for further innovation and advancement.
作者
马祥天
王立志
黄华
MA Xiang-Tian;WANG Li-Zhi;HUANG Hua(School of Computer Science and Technology,Beijing Institute of Technology,Beijing 100081;School of Artificial Intelligence,Beijing Normal University,Beijing 100875)
出处
《计算机学报》
EI
CAS
CSCD
北大核心
2024年第1期190-212,共23页
Chinese Journal of Computers
基金
国家自然科学基金(62131003)资助。
关键词
快照光谱成像
编码孔径
图像重构
优化模型
深度学习
snapshot spectral imaging
coded aperture
image reconstruction
optimization model
deep learning
