单幅RGB图像计算光谱成像的深度学习研究综述

Research of Deep Learning-Based Computational Spectral Imaging for Single RGB Image

为解决传统的光谱成像方法成本高、图像采集时间较长的问题,深度学习被引入计算光谱成像来研究如何从单幅RGB图像中重建光谱,为各种计算机视觉应用提供辅助信息。当前对基于深度学习的单幅RGB图像计算光谱成像方法还未有全面、系统的深入认识与研究。为此针对计算光谱成像所使用的深度学习算法和网络模型进行了系统的归纳、分析和对比。基于CNN(convolutional neural networks)、GAN(generative adversarial networks)、注意力和Transformer四个类别详细梳理了近几年重建性能优异的有监督学习方法;基于自编码器和领域自适应两类别分析、探讨、比较了热度较高的无监督学习方法。同时列举了算法常用数据集和评估指标,对未来的研究趋势和发展方向进行了展望。

Deep learning is introduced into computational spectral imaging to address the high cost and long image acquisition time of traditional spectral imaging methods to investigate how spectral information can be recovered from a single RGB image to provide assistance for various computer vision applications.Computational spectral imaging methods for single RGB images based on deep learning lack comprehensive and systematic research.Deep learning algorithms and network models used for computational spectral imaging are summarised,analyzed and compared.The four categories of CNN(convolutional neural networks),GAN(generative adversarial networks),Attention and Transformer are used to sort out supervised learning methods with excellent reconstruction performance in recent years.The unsupervised learning methods are discussed in terms of both self-encoders and domain adaptation.Datasets and evaluation metrics commonly used for the algorithms are listed,and future research trends and development directions are given.