GA-BP神经网络结合PCA的多基色颜色预测模型

Multi-color Prediction Model Based on BP-NN Optimized by GA and PCA

颜色分区法是简化多基色颜色复制的主要方法,本研究以3基色为组将7基色色空间分成6个子空间,结合主成分分析(PCA)对分区内颜色样本的光谱反射率降维,通过3层BP神经网络,建立样本网点面积率与多基色复制色光谱反射率的转换模型,并通过遗传算法(GA)对神经网络权值阈值进行优化,提高多基色复制的颜色预测精度。实验结果表明,在各分区训练样本数为64、检测样本数为216时,GA-BP神经网络模型颜色预测的平均色差(ΔE*ab)为1.669,光谱均方根误差(RMSE)为0.7%,预测精度和稳定性均高于BP神经网络模型和胞元Neugebauer模型。最后,将训练样本数为64的GA-BP模型与训练样本数量为125,216,343的BP神经网络模型(平均ΔE*ab分别为3.267,2.776,2.175,光谱RMSE为0.97%,0.79%,0.76%)进行了比较,结果表明训练样本数为64的GA-BP模型的预测精度与训练样本数量为343的BP神经网络模型相当。GA-BP模型仅需少量样本即可实现高精度的颜色预测,在应用中具有良好的可移植性。

Color partition method is the main method of multi-color reproduction. In this paper,the whole color area of 7 primary colors was divided into 6 partitions in group of 3 primary colors,and principal component analysis( PCA) was used to dimensionality reduction of the color spectrum reflectance. A three layers BP neural network was established to describe the transformation model of dot area coverage and spectral reflectance. Genetic algorithm( GA) was adopted for optimizing the weight threshold of BP NN to improve the prediction accuracy of model. Experimental results show that GA-BP model has higher prediction precision and stability compared with BP ANN and cell neugebauer model. When the number of training sample is 64 and test sample is 216 in the partition,the optimized model can predict color with the accuracy of 1. 669 mean ΔE*aband 0. 7% spectral RMSE. By comparing with the non-optimized model with the training sample numbers of 125,216,343( mean ΔE*abare 3. 267,2. 776,2. 175 and spectral RMSE are 0. 97%,0. 79%,0. 76%),the prediction accuracy of GA-BP model with the training sample numbers of 64 is equal to the accuracy of the non-optimized model with the training sample number of 343. The results show that GA-BP model can reach high precision color reproduction with small amount of samples,andhave good portability in practice.