Comparative analysis of deep learning models for garment segmentation in textile recycling

Abstract The textile industry is a key player in the global economy, but its expansion and resource consumption raise sustainability concerns. Addressing these issues is critical to fostering a more sustainable future. One key challenge in garment recycling is the removal of accessories and inserts before fibers can be recovered. This study presents a comparative analysis of deep learning models for garment segmentation as an enabling step for future automated textile recycling workflows. Accurate garment segmentation can support subsequent tasks such as garment classification, accessory localization, and selective removal, improving the consistency and efficiency of pre-processing before fiber recovery. Several state-of-the-art models, including You Only Look Once version 8 (YOLOv8), Inception Residual Network version 2 with U-Net (InceptionResNetV2-UNet), U2-Net, and Mask Region-based Convolutional Neural Network (Mask R-CNN), were evaluated to understand their strengths and weaknesses in handling garment complexity. The highest performance was demonstrated by the combination of the YOLOv8n detector plus U2-Net, with 99.56% accuracy, 99.11% F1-score, 98.23% Jaccard Index, 99.91% recall, and 98.33% precision, when tested on 347 laboratory images, showcasing robust segmentation capabilities under controlled conditions. The results suggest that this combination is well-suited for garment segmentation tasks and can provide a useful computer vision foundation for future accessory localization and automated removal modules. This comparison serves as a basis for future studies focused on improving segmentation robustness and integrating garment segmentation with downstream textile recycling processes.