MBA-DNet: A mask block attention-based foreign matter detection network for tobacco packages

Area of Science:

• Computer Vision
• Machine Learning
• Quality Control

Background:

• Detecting foreign matter like varia and mildew in tobacco packaging is crucial for quality control.
• Current object detection algorithms struggle with small targets in tobacco package inspections.

Purpose of the Study:

• To improve the detection of small foreign matter in the cutaway view of tobacco packages.
• To enhance the global modeling capability of object detection networks for this specific application.

Main Methods:

• Proposed Mask Block Attention (MBA) and integrated it into the feature extraction network.
• Developed a K-fold packet slicing defect dataset (K-PSDDS).
• Evaluated the enhanced DINO (DETR with an improved denoising anchor box for end-to-end target detection) + MBA model on K-PSDDS.

Main Results:

• The DINO + MBA model achieved AP50 of 94.9% and APbbox of 47.7%.
• Demonstrated a performance improvement of 0.3% and 0.9% over the baseline DINO network.
• Outperformed other detection algorithms, including fast region-based convolutional neural networks.

Conclusions:

• The proposed MBA significantly enhances object detection capabilities for foreign matter in tobacco packaging.
• The K-PSDDS dataset and MBA-enhanced DINO model offer a robust solution for automated quality inspection.