A ViT-based Method of Stitching Defect Detection for Packaging Bags by Integrating Image Correction and Transfer Learning Solutions

In the automated packaging production line of industrial systems, various stitching defects of packaging bags are inevitable hence the automated detection technology for stitching defects based on digital image processing plays a crucial role in improving the quality and reliability of the packaging production line. This paper proposes a novel stitching defect detection method based on Vision Transformers. image preprocessing procedures such as image correction and cropping recognition are performed to construct the raw stitching defect dataset. The pre-trained ViT_Base_Patch16_224_In21k and MobileViT-XXS networks are proposed by transfer learning methods to improve the recognition of stitching defects. Six sets of experiments are performed on image data taken on a packaging production line, and on the test set, the defect detection models’ accuracy rates achieve 0.95 and 0.989, respectively. The experimental results reveal that the proposed method can be effectively applied for online automatic detection of stitching defects in packaging production lines and has industrial application significance.