Tobacco shred varieties classification using Multi-Scale-X-ResNet network and machine vision.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Science

Background:

• Accurate classification of tobacco shred types is crucial for blending ratios and quality control.
• Distinguishing between similar tobacco types like expanded tobacco silk and tobacco silk is challenging due to macro-scale similarities and irregular shapes.
• Existing machine vision methods face difficulties in recognizing and classifying small, irregularly shaped tobacco shreds.

Purpose of the Study:

• To develop a robust machine vision system for precise classification of four tobacco shred types.
• To propose an optimized deep learning model for enhanced tobacco shred identification.
• To provide an efficient solution for real-time online identification of tobacco types.

Main Methods:

• Image preprocessing using block threshold binarization with optimized parameters.
• Development of a novel deep learning network, MS-X-ResNet (Multi-Scale-X-ResNet), based on ResNet50.
• Integration of multi-scale feature fusion and adjustment of network block structures (A-ResNet, B-ResNet).
• Utilization of the focal loss function to handle class imbalance and improve performance.

Main Results:

• The proposed MS-X-ResNet model achieved a classification accuracy of 96.56% on the tobacco shred dataset.
• Single tobacco shred image recognition was accomplished in 103 ms, demonstrating high efficiency.
• The system successfully addressed challenges posed by small size and irregular shapes of tobacco shreds.

Conclusions:

• The developed image preprocessing and deep learning approach offers a novel solution for tobacco production and quality detection.
• The MS-X-ResNet network provides high accuracy and efficiency for tobacco shred classification.
• This method presents a new pathway for online, real-time type identification applicable to other agricultural products.