RESUMO
Near-infrared spectroscopy (NIRS) has emerged as a key technique for rapid quality detection owing to its fast, non-destructive, and eco-friendly characteristics. However, its practical implementation within the formulation industry is challenging owing to insufficient data, which renders model fitting difficult. The complexity of acquiring spectra and spectral reference values results in limited spectral data, aggravating the problem of low generalization, which diminishes model performance. To address this problem, we introduce what we believe to be a novel approach combining NIRS with Wasserstein generative adversarial networks (WGANs). Specifically, spectral data are collected from representative samples of raw material provided by a formula enterprise. Then, the WGAN augments the database by generating synthetic data resembling the raw spectral data. Finally, we establish various prediction models using the PLSR, SVR, LightGBM, and XGBoost algorithms. Experimental results show the NIRS-WGAN method significantly improves the performance of prediction models, with R2 and RMSE of 0.949 and 1.415 for the chemical components of sugar, respectively, and 0.922 and 0.243 for nicotine. The proposed framework effectively enhances the predictive capabilities of various models, addressing the issue caused by limited training data in NIRS prediction tasks.
RESUMO
Flue-cured tobacco grading plays a crucial role in tobacco leaf purchase and the formulation of tobacco leaf groups. However, the traditional flue-cured tobacco grading mode is usually manual, which is time-consuming, laborious, and subjective. Hence, it is essential to research more efficient and intelligent flue-cured tobacco grading methods. Most existing methods suffer from the more classes less accuracy problem. Meanwhile, limited by different industry applications, the flue-cured tobacco datasets are hard to be obtained publicly. The existing methods employ relatively small and lower resolution tobacco data that are hard to apply in practice. Therefore, aiming at the insufficiency of feature extraction ability and the inadaptability to multiple flue-cured tobacco grades, we collected the largest and highest resolution dataset and proposed an efficient flue-cured tobacco grading method based on deep densely convolutional network (DenseNet). Diverging from other approaches, our method has a unique connectivity pattern of convolutional neural network that concatenates preceding tobacco feature data. This mode connects all previous layers to the subsequent layer directly for tobacco feature transmission. This idea can better extract depth tobacco image information features and transmit each layer's data, thereby reducing the information loss and encouraging tobacco feature reuse. Then, we designed the whole data pre-processing process and experimented with traditional and deep learning algorithms to verify our dataset usability. The experimental results showed that DenseNet could be easily adapted by changing the output of the fully connected layers. With an accuracy of 0.997, significantly higher than the other intelligent tobacco grading methods, DenseNet came to the best model for solving our flue-cured tobacco grading problem.
