Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Supervised and Unsupervised Textile Classification via Near-Infrared Hyperspectral Imaging and Deep Learning

Maria Kainz, Johannes K. Krondorfer, Malte Jaschik, Maria Jernej, Harald Ganster

TL;DR

This work tackles automated textile sorting to enhance recycling by combining near-infrared hyperspectral imaging with deep learning. It evaluates a supervised 1D-CNN for multi-class fiber-type and structure classification and an unsupervised autoencoder for target-textile anomaly detection, assessing generalization under color and structural variations. The study demonstrates robust out-of-sample performance for the CNN and shows that reconstruction-error-based detection in autoencoders can separate non-target textiles, though with limited generalization on blends. The findings highlight the potential for accurate, robust, and scalable textile sorting pipelines that support sustainable recycling, aided by careful preprocessing and dataset design that mimics real-world variability.

Abstract

Recycling textile fibers is critical to reducing the environmental impact of the textile industry. Hyperspectral near-infrared (NIR) imaging combined with advanced deep learning algorithms offers a promising solution for efficient fiber classification and sorting. In this study, we investigate supervised and unsupervised deep learning models and test their generalization capabilities on different textile structures. We show that optimized convolutional neural networks (CNNs) and autoencoder networks achieve robust generalization under varying conditions. These results highlight the potential of hyperspectral imaging and deep learning to advance sustainable textile recycling through accurate and robust classification.

Supervised and Unsupervised Textile Classification via Near-Infrared Hyperspectral Imaging and Deep Learning

TL;DR

This work tackles automated textile sorting to enhance recycling by combining near-infrared hyperspectral imaging with deep learning. It evaluates a supervised 1D-CNN for multi-class fiber-type and structure classification and an unsupervised autoencoder for target-textile anomaly detection, assessing generalization under color and structural variations. The study demonstrates robust out-of-sample performance for the CNN and shows that reconstruction-error-based detection in autoencoders can separate non-target textiles, though with limited generalization on blends. The findings highlight the potential for accurate, robust, and scalable textile sorting pipelines that support sustainable recycling, aided by careful preprocessing and dataset design that mimics real-world variability.

Abstract

Recycling textile fibers is critical to reducing the environmental impact of the textile industry. Hyperspectral near-infrared (NIR) imaging combined with advanced deep learning algorithms offers a promising solution for efficient fiber classification and sorting. In this study, we investigate supervised and unsupervised deep learning models and test their generalization capabilities on different textile structures. We show that optimized convolutional neural networks (CNNs) and autoencoder networks achieve robust generalization under varying conditions. These results highlight the potential of hyperspectral imaging and deep learning to advance sustainable textile recycling through accurate and robust classification.
Paper Structure (17 sections, 3 figures, 2 tables)

This paper contains 17 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Textile data, Acquisition and Preprocessing
  3. Textile categorization:
  4. Hyperspectral imaging:
  5. Dark textiles:
  6. Preprocessing:
  7. Supervised Classification of Hyperspectral Textile Data
  8. Datasets:
  9. Model:
  10. Training:
  11. Results:
  12. Unsupervised Classification Using Autoencoders
  13. Datasets:
  14. Model:
  15. Training:
  16. ...and 2 more sections

Figures (3)

  • Figure 1: Comparison of polyester structure types (Panne Velvet, Satin, Fleece).
  • Figure 2: Confusion matrices for supervised classification of pixel-based (left) and object-based (right) analysis on D2 (top) and D3 (bottom). On the test set of D1 all data are correctly classified. Also on D2 and D3 exceptional performance can be observed, with slight misclassification rates for cotton blends or structurally different cotton fibers. With respect to an object-based accuracy, however, all textiles are classified correctly, despite the different structure and coloring.
  • Figure 3: Distribution of reconstruction error (RE) for different textiles. The decision threshold is based on the 95% quantile of the RE distribution over D4. Pure textiles from D5 can be well distinguished and yield 100% recognition (bottom), whereas cotton blends are much harder to distinguish and only reduced recognition can be observed (upper left). The RE distribution of different cotton types from D6 is much wider than on the training set and only moderate generalization is observed, leading to reduced accuracy. The corresponding pixel- and object-based accuracies are given in Table \ref{['kainz_maria:tab:autoencoder_results']}.