用于奶油色素定量分析的注意力残差网络设计与验证

Design and verification of attention residual network for quantitative analysis of cream pigments

针对样品温度变化问题,由于近红外光谱对温度等物理条件变化十分敏感,以奶油中的靛蓝色素作为光谱定量分析数据,提出了一种变温注意力残差网络解决方案。变温注意力残差网络融合温度以及光谱特征,其主干结构使用并发空间和通道挤压和激励注意力机制对残差块处理后的特征进行整合增强。随后采用最大池化和随机丢弃层进行特征降维和模型正则化。将去掉注意力模块的网络与六种深度学习常用的回归分析网络对比,验证其在领域的高适用性。将变温注意力残差网络与6种网络中最佳模型的3种优化形式对比,验证其高性能。最后对模型调优,训练和测试损失差缩小至0.0005,决定系数和相对分析误差达到了最佳值0.9293和3.7031,表明该模型能在实践中对变温条件下的光谱定量分析。

Temperature change of sample causes fluctuation to its spectrum.As near-infrared spectroscopy is very sensitive to changes in physical conditions such as temperature,we took the indigo pigment in cream as the spectral quantitative analysis data and proposed a variable temperature attention residual network.This network integrates temperature and spectral features,and its backbone structure adopts a concurrent spatial and channel squeeze and excitation attention mechanism to integrate and enhance the features processed by the residual block.Subsequently,we used maximum pooling and random dropout layers for feature dimensionality reduction and model regularization.By comparing the network without the attention module with six commonly used regression analysis networks in deep learning,we verified its high applicability in this field;by comparing the variable temperature attention residual network with three optimization forms of the best model among the six networks,we verified its high performance.After we tuned the model,the difference between the training and test losses was reduced to 0.0005,and the coefficient of determination and the relative analysis error reached the best values of 0.9293 and 3.7031,indicating that the model can perform quantitative analysis of spectra under variable temperature conditions in practice.