TUTNet - A Segmentation Network for Fabric Sewing Thread in the Texture Background

Sewing thread segmentation can help locate the defects in fabric sewing process, which is a challenging problem in factory quality control. In this paper, we propose a twin U-shaped Transformer Network (TUTNet) for sewing thread segmentation. To learn high-dimensional abstract features and capture the sewing direction while ensuring the smoothness of segmentation, we propose a Twin-Attention Module (TAM), which is composed of Sewing Thread Direction (STD) and Sewing Thread Smoothness (STS) blocks. To better combine low-dimensional features with fine-grained features, and generate more continuous results, we propose a Scaled Feature Fusion Module (SFFM), which is used to calculate losses at multiple levels. We also propose a large-scale sewing thread segmentation (STSeg) dataset, which is collected from real clothing production factory. Finally, we conduct experiments on both STSeg and DGRE datasets to verify the effectiveness of our method for sewing thread and other linear object segmentation.