基于遗传算法选择多光源下的光谱反射率重构研究

为解决基于RGB三通道信息值重构光谱反射率精度不理想的问题，提出了一种优化的基于RGB三通道信息的光谱反射率重构算法。首先编码产生随机选择多个光源的个体，RGB三通道值通过多项式回归算法预测多个光源下的三刺激值，并采用伪逆法进行多光源下的光谱反射率重构，然后将样本的重构精度作为个体的适应度评估值，以优胜劣汰，适者生存为原则对个体进行选择、交叉、变异操作，最后得到适用于颜色样本光谱重构的多个光源与基于这些光源重构得到的光谱反射率。实验选用Munsell颜色集作为训练样本集，RC24色卡、SG140色卡作为检测样本集，8个标准光源和82个发光二极管光源作为实验光源，采用该算法从90个光源中选取最优的光源组合并重构得到样本的光谱数据，并与Zhang提出的基于穷举法选择的多光源下的光谱重构方法和A光源下的伪逆法进行了重构精度对比。实验结果显示该研究提出的方法随着光源个数的增加，光谱反射率重构精度提高，特别是光源个数增加到3时，光谱重构精度提高的幅度最大。在三种重构方法中，该方法重构RC24的平均色差和平均光谱均方根误差分别为0.332 4和0.002 9，而Zhang的方法与伪逆法的平均色差分别为0.429 3和3.266，平均光谱均方根误差分别为0.029 7和0.004 8；该文方法重构SG140的平均色差和平均光谱均方根误差分别为0.486 2和0.007 3，而Zhang的方法与伪逆法的平均色差分别为0.544 8和3.821 9，平均光谱均方根误差分别为0.035 6和0.013 3。结果表明基于多光源下的光谱反射率重构精度明显优于基于单个光源下的重构精度，而基于遗传算法的多光源选择方法又优于穷举法，它能够根据颜色样本自动寻找到最优光源组合，从而基于最优多光源下的三刺激值重构样本的光谱反射率，提高了光谱反射率重构的精度。

Research on Spectral Reflectance Reconstruction Based on Genetic Algorithm for Selecting Multi-Illuminants

In order to solve the problem that the accuracy of the spectral reflectance reconstruction based on RGB three-channel values is not ideal, an optimized spectral reflectance reconstruction algorithm based on RGB three-channel information was proposed. Firstly we coded to generate individuals with multiple illuminants selected randomly, and RGB three-channel values were used to predict CIE XYZ values under multi-illuminant by polynomial regression algorithm, and the pseudo-inverse method was used to reconstruct the spectral reflectance. Then the reconstruction accuracy of the sample was taken as the fitness evaluation value of the individual, and the individuals were selected, crossed, and mutated based on the principle of survival of the fittest. Finally, the combination of multiple illuminants which were suitable for spectral reconstruction of color samples were obtained, and then used to reconstruct the spectral of color samples. Munsell color set was used as training samples, RC24 chart and SG140 chart were used as test samples, and 8 standard illuminants and 82 LED light sources were used as experimental light sources. The proposed method was used to select the optimal combination from the 90 illumination sources and reconstruct the spectral reflectance of the test samples, and compared with the method based on multi-illuminant selected by exhaustive method proposed by Zhang and the pseudo-inverse method under A light source. The experimental results show that the spectral reflectance reconstruction accuracy is improved as the number of light sources increases, and the increase achieves the most when the number of light sources reach to 3. Among the three reconstruction methods, the average color difference and average root mean square error of the proposed method are 0.332 4 and 0.002 9 respectively for the RC24 chart，while the average color difference of the Zhang and pseudo-inverse methods are 0.429 3 and 3.266 respectively, and their average root mean square error are 0.029 7 and 0.004 8. For the SG140 chart, the average color difference and average root mean square error of the proposed method are 0.486 2 and 0.007 3 respectively, while the average color differences of the Zhang and pseudo-inverse methods are 0.544 8 and 3.821 9 respectively, and the average root mean errors are 0.035 6 and 0.013 3 respectively. The results show that the spectral reconstruction accuracy obtained under multi-illuminant is obviously superior to the one obtained under a single light source, and the results of the multi-illuminant selection method based on genetic algorithm is better than those of the exhaustive method. The genetic algorithm can automatically find the optimal illuminant combination according to the color samples, so as to reconstruct the spectral reflectance of the sample based on the optimal combination improving the accuracy of spectral reconstruction.