基于人工神经网络的光谱反射率重建

Reconstruction of Spectral Reflectance Based on Artificial Neural Networks

目的研究基于BP神经网络法和FNN神经网络法重构图像光谱反射率的精度。方法以SG标准色卡作为训练样本,分别使用BP和FNN神经网络法,对测试样本DC标准色卡的光谱反射率进行预测,并利用CIEL*a*b*色差公式、均方根误差(ERMS)和光谱匹配精度(GFC)对结果进行评价。结果 BP和FNN神经网络重构的光谱反射率平均色差(ΔEab)分别为2.997和3.071,平均均方根误差(ERMS)分别为0.056和0.049,平均光谱匹配精度(GFC)分别为0.987和0.991。结论 2种神经网络方法重构的光谱反射率具有相当优越的色度和光谱精度。相比于FNN神经网络,BP神经网络更加适合于光谱图像的获取领域。

The aim of this work was to study the accuracy of the image spectral reflectance reconstructed based on BP neural network and FNN neural network. SG standard color card was taken as the training sample to predict the spectral reflectance of DC standard color card using BP neural network and FNN neural network, respectively, and then the results were evaluated and analyzed with CIE L＊a＊b＊color difference, error root mean square and Goodness-Fitting Coefficient. The average color difference, average error root mean and average Goodness-Fitting Coefficient of reflectance reconstructed with BP neural network were 2.997, 0.056, and 0.981, respectively, while those reconstructed with FNN neural network were 3.071, 0.049, and 0.991, respectively. The spectral reflectance reconstructed by both neural networks had good color and spectral accuracy. Compared to the FNN neural network, BP neural network was more suitable for the field of spectral image acquisition.