Research on Spectral Reflectance Reconstruction Methods Based on Multispectral Imaging

Spectral reflectance, as an intrinsic property of an object's surface color, possesses uniqueness and plays a vital role in distinguishing material compositions. However, traditional spectrophotometers lack imaging capabilities, and hyperspectral imaging systems face limitations in widespread application due to their high cost, bulky hardware, and complex data processing requirements. While multispectral imaging offers a lower-cost alternative, its limited spectral resolution hinders accurate differentiation of substances with similar reflectance characteristics. To address the challenge of reconstructing hyperspectral data from multispectral inputs, this paper investigates three representative methods: Convolutional Neural Networks (CNN), Autoencoders, and Dictionary Learning. Experiments are conducted on a custom-built color dataset. Comparative results demonstrate that the CNN-based approach outperforms others in reconstructing spectral reflectance from 17channel multispectral data to 5000 -band hyperspectral data. The proposed method achieves a root mean square error (RMSE) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.03 \pm 0.02$</tex> in the visible to near-infrared range (370-1039nm) and shows superior spectral detail preservation, particularly in critical regions such as the chlorophyll absorption peak (550 nm) and the near-infrared range (760-1000nm).