Identification of textile fiber composition in waste textiles using improved CNN and near-infrared spectroscopy

Accurate classification of waste textiles is a crucial prerequisite for their efficient recycling, especially in the fields of industrial waste processing and textile recycling, where more precise identification and classification are required to achieve effective resource reuse. However, current traditional methods for waste clothing classification are inefficient, not only time-consuming and labor-intensive for manual sorting but also accuracy recognition-lacking, unable to meet the demands of large-scale processing. Furthermore, traditional methods often incur high operational costs, including labor, equipment, and maintenance expenses, thus failing to satisfy the requirements of low cost and high efficiency in the recycling system. This study introduces a novel method for the identification of waste textile materials using near-infrared (NIR) spectroscopy, based on an enhanced Convolutional Neural Network (CNN) model named FabricNet. The collected NIR spectral data were preprocessed using moving average filtering (MAF) for smoothing and normalization. The FabricNet model integrates the ResNet-18 architecture, Squeeze-and-Excitation (SE) attention mechanism, and Dropout regularization techniques. Experimental results demonstrate that, on a dataset comprising 280 fabric samples (including 11 pure textiles, 22 binary blends, and 31 multi-component blends), the validation accuracy of the FabricNet model reached 97.31%, significantly outperforming other common CNN models such as AlexNet, VGGNet, and GoogleNet. Furthermore, when tested on an independent set of 120 textile samples not used in the training, the model achieved recognition precisions exceeding 95% for pure textiles and over 90% for blended textiles. This method provides a viable and effective technological pathway for the automated classification of waste textiles, significantly enhancing their recycling rates.