建立CRT色度变换的神经网络模型

针对7点LOG空间分布方案的343个训练样本,提出了采用2个隐层和少节点网络结构的方案,并用拟牛顿法训练神经网络模型。采用10点LOG空间分布方案中不同于训练样本的216个检验样本实时监控训练过程,以避免出现“过训练”现象,从而求得全局极小点邻域内的可行解,建立从CRT的R,G,B到CIE的X,Y,Z色度空间变换的BP模型。实例计算表明,该模型在收敛性、训练时间和泛化能力等方面均明显优于采用3～4个隐层方案的模型;模型的色差平均转换精度接近0.60个CIELUV色差单位,标准离差为0.57个色差单位,而4个隐层方案模型的色差平均精度和标准离差分别为1.53和0.77个CIELUV色差单位。

Feasible model for CRT color conversion using neural networks

As to the 343 training set data according to the principle with 7 LOG space, a neural network topology with two hidden layers with a few neurons and the efficient and robust quasi-Newton method are applied to establish neural network model in this paper. The 216 verification set data with 10 LOG space different from the training set data are used to monitor the training process simultaneously to escape from local minimum. The color notation conversion model between the RGB space of the CRT and the XYZ space of CIE system was established using neural networks. The case study shows that the converging speed, the training time and the generalization of the model with two hidden layers and a few neurons are better than that of models with 3 or 4 hidden layers established in the past. The average precision of the color notation conversion of the model established in this paper is about 0.60CIELUV units, and the standard deviation 0.57, in contrast, they were 1.53 and 0.77 of the model with 4 hidden layers established in the past.