ABSTRACT

Accurately estimating spectral reflectance functions from color camera images is a hot research subject that demonstrates tremendous potential for illuminating engineering and computer vision applications. However, the impact of the illumination spectrum and camera responsivity (system functions) on estimation accuracy has not been systematically studied so far, nor has the impact of a “training” spectral reflectance set. In this study, a dual imaging reflectance optimization system is used based on a neural network and optimal system functions that are respectively trained and optimized using several sample sets. Simulations showed that such optimal systems, trained and optimized with the IES TM-30 spectral reflectance set, can have a substantially higher estimation accuracy compared to “real” systems composed of commercially available projector spectra and camera responsivities and that they are sufficiently robust under small changes in system function peak wavelength and spectral width due to changes in working temperature or with passing time. An analysis of the impact of the specific sample set database adopted for neural net training on estimation accuracy showed that training with the IES set results in good and stable performance, even for other sample sets and different illumination spectra. Training with the spectrally uniform IES spectral reflectance set is therefore advised for general-purpose, high-accuracy reflectance estimation systems. A comparison with a state-of-the-art method shows that the proposed method has a higher color prediction accuracy and a significantly shorter running time for realistic images with high resolution.

摘要

从彩色相机图像准确估计光谱反射函数是一个热门的研究课题，它在照明工程和计算机视觉应用中展示了巨大的潜力。然而，照明光谱和相机响应度（系统函数）对估计精度的影响到目前为止还没有被系统地研究过，“训练”光谱反射集的影响也没有。在这项研究中，使用基于神经网络和优化系统函数的双成像反射优化系统，这些系统函数分别使用多个样本集进行训练和优化。模拟表明，使用 IES TM-30 光谱反射集训练和优化的此类最佳系统，与由商用投影仪光谱和相机响应度组成的“真实”系统相比，可以具有更高的估计精度，并且它们在系统功能峰值波长和光谱宽度由于工作温度变化或随着时间的变化而产生的微小变化下具有足够的鲁棒性。分析神经网络训练所采用的特定样本集数据库对估计精度的影响表明，使用 IES 集进行训练，即使对于其他样本集和不同的照明光谱，也能获得良好且稳定的性能。因此，建议针对通用、高精度反射率估计系统使用光谱均匀的 IES 光谱反射率集进行训练。