光谱一般是指波长覆盖紫外到红外波段的电磁波频谱，包含了电磁波与物质相互作用的丰富信息。每种物质都有特征的吸收、发射或散射光谱，因而光谱也被称为物质的指纹，可以用于物质的鉴别和分析。光谱成像技术能够获取成像视场内各点的光谱信息，在遥感测绘、精准农业、疾病诊疗、机器视觉等诸多领域有着巨大的应用需求。基于空间扫描或波长扫描的传统光谱成像设备体积庞大，无法获取动态的光谱信息。本论文以实现高空间分辨率、高光谱分辨率的实时光谱成像芯片为研究目标，基于超表面宽带调制结合计算重建的光谱成像方案，围绕提升光谱重建性能、降低角度敏感性和提高光谱成像速度等关键问题展开研究，主要成果和创新点如下：提出了一种自由形状超表面的正向设计方法，其结构设计自由度相比于规则形状提升了2~3个数量级，有效降低了超表面单元光谱调制函数间的相关性以提升光谱分辨率。在实验上实现了工作波段450~750nm、光谱分辨率0.5nm、空间分辨率360×440的光谱成像芯片。在此基础上，进一步基于形状编码实现了自由形状超表面的逆向设计，利用神经网络实现仿真光谱重建进行性能评估，其抗噪声性能相比正向设计结果提升了约1倍。提出利用金属-介质复合型超表面中的局域表面等离子体共振模式和金属-绝缘体-金属法布里-珀罗腔模式，降低超表面单元的角度敏感性。结合自由形状超表面，利用遗传算法优化设计了一组角度不敏感的金超表面单元，通过神经网络进行仿真光谱重建，验证了角度不敏感的光谱成像。进一步制备并测试了角度不敏感的光谱成像芯片，实现了30°视场角下似然度超过98%的角度不敏感的光谱重建，并演示了对于标准色卡的光谱成像。提出利用深度展开神经网络实现光谱图像的快速重建，解决了传统逐点迭代光谱重建算法耗时长和重建图像存在马赛克现象的问题。利用开源光谱图像数据集进行网络训练，并利用高斯噪声模拟测量噪声，实现了高精度的毫秒级光谱图像重建，重建速度相比于传统的逐点迭代重建算法提升了约5个数量级。演示了实时光谱视频成像，有望解决自动驾驶场景中的同色异谱识别问题，展示出实时光谱成像芯片在机器视觉领域的应用潜力。

The spectrum generally refers to the electromagnetic spectrum covering the wavelength from ultraviolet to infrared, which contains rich information about the interaction between electromagnetic wave and matter. Each substance has a characteristic absorption, emission or scattering spectrum, so the spectrum is also called the fingerprint of the substance, which can be used for the identification and analysis of the substance. Spectral imaging technology can obtain the spectral information of each point in the imaging field of view, which has a huge application demand in many fields such as remote sensing, precision agriculture, disease diagnosis and machine vision. Traditional spectral imaging devices based on spatial scanning or wavelength scanning are bulky and cannot obtain dynamic spectral information. Based on the spectral imaging scheme using broadband modulation of metasurfaces with computational reconstruction, this paper aims to realize a real-time spectral imaging chip with both high spatial resolution and high spectral resolution. The research focuses on key issues including improving spectral reconstruction performance, reducing angle sensitivity and improving spectral imaging speed. The main achievements and innovation points are summarized as follows:A forward design method of freeform shaped metasurface is proposed. The degree of freedom of structural design is improved by 2~3 orders of magnitude compared with regular shape, which effectively reduces the correlation between spectral modulation functions of metasurface units to improve spectral resolution. Experimentally, a spectral imaging chip with a working band of 450~750 nm, a spectral resolution of 0.5 nm and a spatial resolution of 360×440 was realized. Based on that, the inverse design of freeform shaped metasurface was realized based on shape coding, and the anti-noise ability is further improved by about 1 times compared with the forward design using the neural network to implement the simulated spectral reconstruction for performance evaluation.The localized surface plasmon resonance mode and the metal-insulator-metal Fabry-Perot cavity mode in the metal-dielectric composite metasurface are proposed to reduce the angle sensitivity of the metasurface units. Combined with the freeform shaped metasurface, a group of angle-insensitive gold metasurface units was optimized by genetic algorithm. The angle-insensitive spectral imaging was verified using the neural network to realize simulated spectral reconstruction. Experimentally, an angle-insensitive spectral imaging chip was fabricated and tested, and angle-insensitive spectral reconstruction with a fidelity of over 98% at a 30-degree field of view was achieved. Besides, the spectral imaging for a standard color board was demonstrated.A deep unrolled neural network is proposed to realize rapid reconstruction of spectral images, which solves the problem that traditional point-by-point iterative spectral reconstruction algorithm is time-consuming and the reconstructed image has mosaic phenomenon. The open source spectral image datasets are used for network training, and the Gaussian noise is used as simulated measurement noise. The high-precision millisecon-level spectral image reconstruction was realized with the reconstruction speed improved by about 5 orders of magnitude compared with the traditional point-by-point iterative reconstruction algorithm. Real-time spectral video imaging was demonstrated, which is expected to solve the problem of metamerism recognition in autonomous driving scenes, showing the application potential of real-time spectral imaging chips in the field of machine vision.