StitchingNet-Seg: a semantic segmentation dataset of industrial sewing stitch defects

Abstract

Adoption of automated quality control in the textile and apparel industry has lagged owing to inherent features, such as labor-intensive tasks and deformability of flexible fabrics. Precise detection and localized diagnosis of sewing stitch defects remain critical challenges because existing datasets often provide only image-level labels or focus exclusively on raw fabric anomalies. Therefore, we propose StitchingNet-Seg, a comprehensive, large-scale semantic segmentation dataset designed for industrial sewing processes. Derived from the StitchingNet dataset, this dataset comprises 10,836 images covering 11 representative fabric types, 7 defect classes and 1 normal class (8 classes in total). Each image is annotated with pixel-level masks to capture intricate geometric features. To validate the usability and technical quality of the dataset, benchmark experiments are conducted using representative convolutional neural network-based architectures and vision transformer-based models. The evaluation results demonstrated robust overall learnability across diverse architectures. However, they also revealed relatively lower segmentation performance on specific defect classes characterized by subtle thread tension variations and 3D structural deformations. This confirms that StitchingNet-Seg provides a reliable foundation for developing intelligent automated optical inspection systems for the global garment industry.