Semantics-Adaptive Multitask Defect Detection in Vacuum Packaging

Abstract

Vacuum packaging effectively prevents food contamination and facilitates portability and storage. Visual detection technology can rapidly identify packaging issues and ensure food quality and is a crucial method to address food safety concerns. However, due to the extensive semantic interference and imbalanced distribution of defects in packaging images, defect detection faces significant challenges. In this article, a semantic-adaptive multitask defect detection approach for vacuum packaging is proposed to handle these problems. First, a vacuum packaging dataset is created by capturing packaging samples on the production line using an industrial camera. Next, addressing the issue of semantic interference, a novel adaptive region partitioning method, integrating U-net and image-processing techniques, is proposed to enhance the identification of defects across different regions. Finally, considering the imbalance in defect distribution, the disentangled representations of abnormalities (DRA) model with diverse modes is employed for defect detection tasks across multiple regions in this article. The introduction of a convolutional block attention module (CBAM) in the DRA network enhances the capability to detect authentic defects. Experimental results show that the method achieves an AUC of 99.47% and an accuracy of 97.82% on the constructed dataset, showcasing significant improvement over alternative methods. In conclusion, this study not only enhances the defect detection capabilities of food vacuum packaging but also provides solutions to common issues such as semantic interference and imbalanced distribution of defects.