Abstract:

In recent years, the reconstruction of the three-dimensional (3D) temperature field from the flame light-field image by the method of deep learning has been a new direction in radiation thermometry. The traditional temperature field reconstruction network still uses the feature extraction method of the plane image, which not only ignores the 3D ray information recorded in the light-field image but also fails to consider the classification of tracing rays. Therefore, the absence of prior information and the hybridity of different types of features have a negative impact on the reconstruction accuracy of temperature field. This article optimized the feature extraction process in the network. Firstly, the view angle information of the sub-aperture image was added to the network input. Then the spatial and angular features of the light-field were extracted by the double-branch convolution method. Finally, the attention mechanism was used to model the importance of features at different scales. The effectiveness of the above factors on the reconstruction accuracy was verified by orthogonal experiments. The simulation results showed that the Mean Relative Error (MRE) of the temperature field reconstructed by the optimization method was reduced by 44.82%, and the Maximum Relative Error (MMRE) was reduced by 34.76% compared to traditional networks.

Key words: temperature field reconstruction, optimization, neural networks, numerical analysis

摘要：

近年来，利用深度学习的方法从光场火焰图像中重建三维温度分布已成为辐射测温研究领域的一个新方向。传统的温度场重建网络仍沿用平面图像的特征提取方法，不仅忽略了光场图像中记录的三维光线信息，而且没有充分考虑追迹光线的分类问题，致使先验信息的缺失和不同类型特征的杂糅，对温度场的重建精度产生负面影响。本文对网络中的特征提取过程进行了优化，首先将子孔径图像的视角信息添加到网络输入端，然后使用双分支卷积方法分别提取光场的空间特征和角度特征，最后利用注意力机制对特征在不同尺度上的重要性进行建模。本文通过正交实验验证了上述因素对重建精度的影响。仿真结果表明，与传统网络相比，优化方法重建的温度场平均相对误差降低了44.82%，最大相对误差降低了34.76%。

关键词: 温度场重建, 优化, 神经网络, 数值分析