StitchingNet and deep transfer learning method for sewing stitch defect detection

Abstract The clothing manufacturing industry has been slow to adopt new technologies due to its labor-intensive nature, resulting in persistent defect rates averaging around 10%. Deep learning models offer a promising solution for automated sewing defect detection, but their effectiveness relies heavily on large training datasets. This study addresses this challenge by introducing StitchingNet, a comprehensive dataset comprising over 14.5K images of sewing stitches categorized into normal and ten different defect classes. The dataset covers diverse sewing conditions by incorporating eleven fabric-thread combinations. Furthermore, we propose a novel deep transfer learning method that leverages pre-trained deep learning models, including both convolutional neural networks and vision transformers. This method employs a two-stage approach: wide-screening to identify promising candidate models, followed by systematic fine-tuning to optimize their performance for sewing defect detection. Implementing the deep transfer learning method with StitchingNet and evaluating it through on-site feasibility testing and Grad-CAM visualization, we demonstrate that the fine-tuned MobileNetV1-S.C model achieves exceptional performance with an F1-score of 0.99. This highlights the effectiveness of our proposed dataset and deep transfer learning method in developing high-performing models for sewing defect detection, achieving superior results in terms of both accuracy (F1-score) and speed (inference time).