如何在给定照明条件和观测条件的情况下, 由彩色数字相机的响应值重建物体表面光谱反射率, 仍是颜色科学与工程领域一个尚待解决的重要课题。 文章使用奇异值分解的方法将光谱反射率近似为若干基向量的线性组合, 求得组合系数, 然后使用相机输出数据与组合系数训练人工神经网络, 使之能够准确的模拟相机输出与组合系数之间的非线性关系, 最后采用经训练的神经网络, 与基向量结合, 由相机输出准确的重建物体表面的光谱反射率。 实验结果显示, 与线性近似的方法相比, 使用该方法对标准Munsell色块进行反射率重建, 重建误差减小了约67%, 具有高精度、 易实现、 易操作的特点, 可用于对重建精度要求较高的诸多领域。

It is still challenging to reconstruct the spectral reflectance of a surface using digital cameras under given luminance and observation conditions. A new approach to solving the problem which is based on neural network and basis vectors is proposed. At first, the spectral reflectance of the sample surface is measured by spectrometer and the response of an digital camera is recorded. Then the reflectance is represented as a linear combination of several basis vectors by singular value decomposition (SVD). After that, a neural network is trained so that it is able to approximate the relationship between the camera responses and the coefficients of basis vectors accurately. In the end, the spectral reflectance can be reconstructed based on the neural network and basis vectors.In the present paper, the authors reconstructed the spectrum reflectance based on neural network and basis vectors. Compared with other traditional methods, neural network expands the space of unknown function F(S) from linear functions to more general nonlinear functions, which gives more accurate estimation of the coefficients ak and better reflectance reconstruction. Results show that the reflectance of standard Munsell color patch (Matte)can be reconstructed successfully with mean of RMS which is 0.023 4. Compared with linear approximation method, reconstruction of standard Munsell color patch (Matte)using this approach reduces the reconstruction error by 67%. Since the neural network can be implemented by Matlab neural network toolbox, this method can be easily adopted in many other cases. Therefore we conclude that this approach has advantages of higher accuracy, easy implementation and adaptation, thus can be used in many applications.