A New Efficient Multi-Object Detection and Size Calculation for Blended Tobacco Shreds Using an Improved YOLOv7...

Area of Science:

• Computer Vision and Machine Learning
• Industrial Automation and Quality Control

Background:

• Accurate detection of tobacco shreds and calculation of unbroken rates are crucial for cigarette quality inspection.
• Existing machine vision methods struggle with classifying tiny, morphologically complex, and overlapped tobacco shreds.

Purpose of the Study:

• To develop an efficient multi-object detection model for blended tobacco shreds.
• To accurately calculate the unbroken tobacco shred rate using a novel 2D size calculation algorithm.

Main Methods:

• An improved YOLOv7-tiny model with a lightweight Resnet19 backbone was developed.
• Key components like SPPCSPC and the detection head were replaced with SPPFCSPC and a decoupled joint detection head.
• A novel Length-Width Calculation (LWC) algorithm was proposed for 2D size determination of individual shreds.

Main Results:

• The model achieved high detection precision (0.883) and mAP (0.932@.5).
• Testing time was significantly reduced to 4.12 ms.
• Average length and width detection accuracy reached -1.7% and 13.2%, respectively, aligning with manual inspection standards.

Conclusions:

• The developed model offers an efficient solution for multi-object detection and size calculation of blended tobacco shreds in cigarette production.
• This approach provides a robust method for enhancing automated quality inspection in the tobacco industry and similar blended image analysis tasks.