Abstract Growth in the textile industry has directly increased waste production, intensifying environmental problems. Fabric sorting is essential for effective recycling and reducing environmental damage. Traditional fabric-sorting methods are labor-intensive, slow, and prone to errors, limiting their scalability. Current systems mainly focus on fiber composition. However, the structure of the fabric, specifically whether it is knitted or woven, significantly affects its mechanical recyclability, fiber recovery efficiency, and environmental impact. To tackle this issue, we propose using a computer vision approach with deep learning for binary fabric classification (woven vs. knit). This will employ real-world images captured by high-resolution scanners and mobile devices. Instead of a typical fine-tuning procedure, we use a dynamic strategy that adapts based on the model’s architecture. This method aligns the model’s parameters through supervised, instruction-based optimization and contextual calibration suited to each model’s strengths and task needs. We modified the architecture of a DenseNet-121 model by implementing progressive unfreezing, custom classification heads, and architecture-specific optimization strategies. This resulted in a strong, cost-efficient, and stable model with 97.34% validation accuracy and only 0.3M trainable parameters. By using structure-based sorting at scale, this method contributes to more efficient structure-sorting, supporting the broader goal of sustainable textiles. Beyond textiles, this methodology could inspire structure-aware approaches in other recycling domains where material construction plays a crucial role. Graphical Abstract
Huang et al. (Wed,) studied this question.