基于改进CycleGAN的玉米蛋白质含量近红外光谱检测方法

Near Infrared Spectral Detection Method for Maize Protein Content Based on Improved CycleGAN

摘要:

玉米富含的蛋白质对人类健康具有重要意义，精准测定其蛋白质含量是实现玉米优质品种选育与营养含量评估的基础。利用近红外光谱技术可以实现对玉米蛋白质含量无损检测，然而，近红外光谱仪器及生化指标检测成本高昂，机器学习在特征提取方面又存在不足。为此，本文提出了一种基于改进循环生成对抗网络(Cycle-constraint generative adversarial network,CycleGAN)的玉米蛋白质含量近红外光谱检测方法。使用基于CycleGAN的数据降噪模型(NIRDM_CycleGAN)对生成的数据进一步降噪，将处理后的数据用于后续实验。NIRDM_CycleGAN模型引入特征增强模块(ISE)，以增强其降噪能力。在NIRDM_CycleGAN模型降噪处理后，采用中值滤波抑制光谱数据末端的异常峰值。实验结果表明，在完成数据增强后，采用NIRDM_CycleGAN模型对样本进行降噪可进一步提高预测模型性能。本文构建的预测模型(NIRPM_CNN)取得最佳预测结果，预测均方根误差(Root mean square error of prediction,RMSEP)、预测集决定系数(R2p)、相对分析误差(Relative percent deviation,RPD)分别为0.0321、0.9802和7.1183。针对近红外光谱数据样本不足的情况，本文提出的方法可对原始近红外光谱数据进行数据增强与降噪处理，有效扩充数据集规模，从而显著降低检测成本并缩短检测时间，实现了低成本、快速且精准的玉米蛋白质含量近红外光谱无损检测目标。

Abstract:

The abundant protein in maize is important to human health, and accurate determination of protein content is fundamental to variety breeding and nutritional evaluation.Near-infrared spectroscopy (NIRS) was used for the non-destructive measurement of maize protein content.However, NIRS instruments and biochemical reference assays were costly.In addition, the feature-extraction capability of conventional machine-learning methods was limited.Accordingly, a near-infrared spectroscopic method for maize protein determination based on an improved cycle-constraint adversarial network (CycleGAN) was proposed.Unlike most studies centred on generation tasks, a CycleGAN-based data-denoising model (NIRDM _ CycleGAN) was innovatively adopted to further denoise the generated spectra, and the processed data were used in subsequent experiments.A feature-enhancement module ( ISE ) was introduced into the NIRDM_CycleGAN model to strengthen its denoising capability.After denoising by NIRDM_ CycleGAN, a median filter was applied at the spectral end to suppress abnormal peaks.Experimental results showed that after data augmentation, prediction performance was further improved by denoising with the NIRDM_CycleGAN model.Comparative trials demonstrated that the prediction model developed ( NIRPM _ CNN) exhibited the best performance, delivering a root mean square error of prediction (RMSEP) of 0.032 1, a coefficient of determination (R2p ) of 0.980 2, and a relative percent deviation (RPD) of 7.118 3 on the test set.To address the limited NIR sample size, the proposed approach was applied separately for data augmentation and denoising of the raw spectra.The dataset was expanded.As a result, detection cost was reduced and testing time was shortened.Thus low-cost, rapid, and accurate non-destructive NIR detection was achieved.