Sun Xiao-gang, Yuan Gui-bin, Dai Jing-min
Harbin Institute of Technology, Harbin 150001, China.
Guang Pu Xue Yu Guang Pu Fen Xi. 2007 Feb;27(2):213-6.
Considering some defects of back-propagation neural network (BP), a new algorithm combining genetic algorithm (GA) with BP was described. The application of GA-BP to the data processing of multi-spectral thermometry was proposed. The simulation experiments, based on GA-BP algorithm and BP neural network respectively, show that the recognition precision of trained emissivity samples is +/-5 K and +/-10 K respectively, and that of untrained emissivity samples is +/-10 K and +/-20 K respectively. No matter GA-BP algorithm or BP neural network is used, in general, the recognition precision of trained emissivity samples is higher than that of untrained emissivity samples. The recognition precision of true temperature is lower near the edge of sample sets. The GA-BP algorithm was more efficient than the BP neural network in the true temperature measurement.
考虑到反向传播神经网络（BP）的一些缺陷，描述了一种将遗传算法（GA）与BP相结合的新算法。提出了GA - BP算法在多光谱测温数据处理中的应用。分别基于GA - BP算法和BP神经网络进行的仿真实验表明，训练后的发射率样本的识别精度分别为±5 K和±10 K，未训练的发射率样本的识别精度分别为±10 K和±20 K。无论使用GA - BP算法还是BP神经网络，一般来说，训练后的发射率样本的识别精度都高于未训练的发射率样本。在样本集边缘附近，真实温度的识别精度较低。在真实温度测量中，GA - BP算法比BP神经网络更有效。
