Non-destructive detection and classification of textile fibres based on hyperspectral imaging and 1D-CNN
Jiadong Huang,Hongyuan He,Rulin Lv,Guangteng Zhang,Zongxian Zhou,Xiaobin Wang
DOI: https://doi.org/10.1016/j.aca.2022.340238
2022-09-01
Abstract:Textile fibre is very common in daily life, and its classification and identification play an important role in textile recycling, archaeology, public security, and other industries. However, traditional identification methods are time-consuming, laborious, and often destructive to the samples. In order to quickly, accurately, and nondestructively classify and recognize textile fibres, this study established a textile fibre classification and recognition method based on hyperspectral imaging (HSI) and a one-dimensional convolutional neural network (1D-CNN) model. Hyperspectral images of 25 kinds of commercial textile fibres were collected and denoised by pixel fusion. Four traditional machine learning classification models, k-nearest neighbors (KNN), support vector machine (SVM), random forest (RF), and partial least squares-discriminant analysis (PLS-DA), were used to identify the data. The results show that RF has the highest classification accuracy, reaching 91.4%. Then a back propagation neural network (BPNN) model and a one-dimensional convolutional neural network (1D-CNN) model were constructed and compared with the traditional machine learning methods. The results show that the 1D-CNN models have 97.9% and 98.6% accuracy on the training and test sets, respectively. The precision (Pr), sensitivity (Se), specificity (Sp), and F1 score (F1 score) of the models reached 98.7%, 98.6%, 99.9%, and 98.6%, respectively, which were significantly better than the four traditional machine learning models. It seems that 1D-CNN combined with the HSI technique may be a potential method in the detection and classification of textile fibres.