LAUSR

Quantitative analysis of blended fabrics plays a vital role in quality inspection and trade. As a method yielding non-destructive, pollution-free, and easy-to-obtain data, near-infrared spectroscopy is becoming an important supplement to the detection methods of blended fabrics. In other near-infrared spectroscopy analysis tasks, convolutional neural network is becoming a mainstream model. This study aims to further solve the problems of nonlinearity, sample imbalance, and inaccurate labels in the actual near-infrared spectral dataset of blended fabrics. We first proposed to introduce the convolutional network widely applied in similar tasks to establish a quantitative analysis model of the near-infrared spectroscopy of blended fabrics. Two model variants closely related to the spectral data characteristics are considered: a deeper model and a model with feature attention mechanism. For the sample imbalance problem, a label balancing method based on kernel density estimation is proposed, and for the label noise problem, a noise tolerance method based on absolute error truncation is proposed. Comparative experiments are conducted on eight common blended fabric datasets. The results show that the modeling effect of the convolutional neural network model is generally better than that of traditional machine learning methods, especially the network model that can extract more complex time series data features. Thus, the improvements proposed for the data problem can significantly improve the modeling effect in relevant scenarios. The proposed model and methods meet the needs for the non-destructive quantitative analysis of blended fabrics.