Discrimination of dye classes in polyester and cotton textiles via hyperspectral imaging combined with deep-learning-based modeling

With the increasing diversification of the textile market, the rapid and accurate identification of dye classes—specifically azo, anthraquinone, and phthalocyanine—holds significant practical value for product safety, environmental monitoring, and quality control. This study evaluates the application of hyperspectral imaging (HSI) as a non-destructive analytical tool for the discrimination of these dyes on complex textile matrices, including polyester and cotton. To mitigate noise and scattering interference inherent in spectral data, five preprocessing techniques, such as multiple scatter correction (MSC) and standard normal variate (SNV), were systematically investigated. Rather than relying on full-spectrum analysis, a refined feature-selection strategy was implemented. By integrating competitive adaptive reweighted sampling (CARS) with the successive projections algorithm (SPA), spectral dimensionality was significantly reduced, distilling the most representative molecular signatures. Subsequently, a convolutional neural network (CNN) was employed to classify the dye categories. Experimental results demonstrated that the CNN model combined with MSC–CARS preprocessing achieved an overall classification accuracy of 100% on independent sample-level validation sets for both fabric types under controlled experimental conditions. Comparative analysis with support vector machine (SVM) models revealed that the deep-learning-based approach demonstrated improved classification stability. This research establishes a robust analytical protocol and a theoretical foundation for the automated, intelligent inspection of textile chemicals in industrial settings.