中文摘要：

针对多光谱辐射测温理论，建立了一种新的激活函数可调的组合神经网络发射率模型（AFT-CNNE模型）用于从辐射能量中分离出发射率和真温，该模型的计算结果与目标函数的选取有很大的关系。选取了几种不同的目标函数，利用分步式最速下降训练算法对AFT-CNNE进行求解。通过仿真实验，得出了绝对值型目标函数对该模型具有比较快的逼近速度和较高的计算精度的结论，并从统计学的角度证明了采用该目标函数的AFT-CNNE模型是实际发射率模型的无偏估计。该性能优良的目标函数的采用为实现发射率的在线测量奠定了基础。

英文摘要：

A new activation-function-tunable neural network （AFT-CNNE） was constructed towards the multi-spectral thermometry, which could separate emissivity and temperature from target＇s radiation information. The model＇s algorithmic results have relation to the selection of the target function of the model. Several different target functions were discussed. Multistep searching method was used to train the AFT-CNNE model. Results indicate that the absolute form target function has a better astringency speed and precision. This paper proves that the AFT-CNNE model which adopts this target function is the unbiased estimate of real emissivity model, from the point of statistics theory. The adoption of this high-performance target function can be the foundation of realizing the on-line measurement of target＇s real temperature.