Progress in Research on Tobacco Online Inspection Technology Based on Machine Vision

Fig. 10. Microscopic characteristics of cut tobacco^[65]. (a) Leaf filament; (b) stem filamen; (c) expanded leaf filament; (d) reconstituted tobacco leaf filament

Download full size

Fig. 11. Improvement of convolutional neural network model based on LeNet-5^[66]

Download full size

Fig. 12. Improved YOLOv3 network structure^[29]

Download full size

Fig. 13. SAE network structure^[71]

Download full size

Fig. 14. The process of cigarette box recognition system^[72]

Download full size

Fig. 15. Tobacco disease recognition model based on InceptionV3^[74]

Download full size

Fig. 16. ResNet 20 network structure^[75]

Download full size

Fig. 17. The process of transfer learning^[76]

Download full size

Table 1. Comparative analysis of common tobacco product online inspection systems

View table

Table 1. Comparative analysis of common tobacco product online inspection systems

Technical indicator	Tobacco filament defect detection system	Cigarette defect detection system	Cigarette packaging inspection system
Detection targets	Tobacco filament mixture	Multiple cigarettes	Cigarette packaging
Processing speed	Relatively fast	Relatively fast	Fast
Detection accuracy	Average	High	High
Removal difficulty	Complex	Simple	Simple
Technological maturity	Relatively mature	Relatively mature	Relatively mature
Main difficulties	It is difficult to visually identify similar impurities in the tobacco filament，and the removal of impurities is complex，resulting in low efficiency	The cigarette’s surface damage is complex，and there are situations such as empty heads and missing cigarettes，which require high precision from the recognition detection algorithm	The defects in cigarette packaging are diverse，and the production speed of cigarette packs is high，which demands high real-time performance and processing speed from the hardware

Table 2. Comparative analysis of different visual inspection algorithms

View table

Table 2. Comparative analysis of different visual inspection algorithms

Technical Indicator	Based on color space	Based on shape features	Based on texture information	Based on spectral analysis
Methods	Using color histograms and color feature analysis to enhance image contrast	Using edge detection，contour extraction，and morphological processing methods for target detection	Using texture features and template matching to extract target features	Using a spectral camera to obtain target spectral feature information
Features	Intuitive and simple，easy to implement	Sensitive to overall shape changes	Sensitive to local texture features	Analyzes the spectral characteristics of samples
Advantages	Rich color information，sensitive to color differences	Strong detection ability for overall shape defects	capture specific texture features	analyze tobacco composition and quality
Limitations	Greatly affected by lighting conditions and environment	Limited detection ability for local defects	Difficulty in recognizing complex textures	Requires professional equipment，high cost
Applicable Scenarios	Detection of foreign bodies in tobacco filaments，empty heads of cigarettes	Detection of defects and damage in cigarette packaging	Detection of surface damage and stains on cigarettes	Tobacco composition analysis
Technological Maturity	Relatively mature	Relatively mature	Relatively mature	Requires professional equipment and technical support
Complexity	Low	Medium	Medium	High

Table 3. Different defect target detection algorithms based on deep learning

View table

Table 3. Different defect target detection algorithms based on deep learning

Defect target	Tobacco filament component detection	Cigarette box defect detection	Tobacco disease recognition
Common Networks	LeNet-5，Light-VGG Network，AlexNet，GoogleNet，etc.	Faster R-CNN，DarkNet-53，Optimized YOLOv3 Network，Self-encoding Network，etc	Improved-AlexNet，GoogLeNet Inception v2，InceptionV3 Network，ResNet Network，VGG-16 Network，etc.
Methods	Analysis of tobacco filament features，construction of tobacco filament composition classification recognition model	Visual area detection of cigarette boxes	Extraction of disease feature information to build a disease recognition model
Difficulties	Small tobacco filament morphology，numerous impurities	Randomness in the placement angle of cigarette boxes，dense placement of cigarette box images	Tobacco disease symptoms are not obvious in the early stages and some disease features are similar
Improvement Methods	Use of global average pooling to suppress model overfitting	Geometric regularization of cigarette box images，simulation enhancement methods for cigarette box images	Increase the dataset and apply data augmentation techniques；employ ensemble learning，to enhance the model’s robustness and accuracy

Tools

Get Citation

Copy Citation Text

Yusheng Wu, Anhu Li, Yaming Wan, Tianchen Meng. Progress in Research on Tobacco Online Inspection Technology Based on Machine Vision[J]. Laser & Optoelectronics Progress, 2024, 61(8): 0800003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Reviews

Received: May. 18, 2023

Accepted: Jun. 20, 2023

Published Online: Mar. 5, 2024

The Author Email: Yusheng Wu (21480276@qq.com), Anhu Li (lah@tongji.edu.cn)

DOI:10.3788/LOP231332

CSTR:32186.14.LOP231332

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology