Constructing a Xiangyunsha Defect Dataset and a Lightweight Detection Model for Sustainability

Abstract

Xiangyunsha (Gambiered Guangdong silk) is a naturally dyed, handcrafted textile whose sustainable value is constrained by material waste arising from conventional defect-rejection practices. This study reconceptualizes conventionally defined defects, particularly crackle irregularities and stains, as potential design resources rather than production failures. Through field-based image acquisition and on-site observation in three heritage workshops in Foshan, China, we constructed a specialized Xiangyunsha defect-image dataset comprising 1831 annotated images. In the revised experimental protocol, the dataset was divided into 1400 training, 331 validation, and 100 held-out test images, with all model development experiments conducted using only the training and validation subsets. We compared a conventional five-class taxonomy with a task-oriented binary scheme and evaluated multiple YOLO26 model scales, architectural modifications, and two representative lightweight YOLO detectors. Based exclusively on the validation results across three predefined random seeds, YOLO26s-DWR + DynamicConv was selected as the final modified architecture. Across the three independently trained runs corresponding to seeds 0, 42, and 123, YOLO26s-DWR + DynamicConv achieved a mean held-out test mAP@0.5 of 46.1 ± 1.8%, compared with 39.8 ± 4.7% for the YOLO26s baseline. Beyond detection, we further conducted a descriptive design case study in which localized Xiangyunsha surface irregularities informed human-led design interventions, including targeted artistic printing/dyeing for textile_abnormal regions and localized pattern cutting for Stain_Patch regions. The resulting detection–interpretation–design intervention workflow positions AI as a decision support tool for artisans and designers rather than as an autonomous quality control system. This study demonstrates the feasibility of linking defect localization with human-led material-retention strategies, while quantitative waste reduction, external generalizability, and lifecycle benefits remain to be established in future studies.