A quantitative detection method for maize kernel broken rate based on the optimisation of the MSA transformer...

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Mechanized harvesting of maize kernels leads to breakage, impacting food quality and profitability.
• Existing detection methods for kernel breakage are qualitative and lack precision.
• Accurate quantitative assessment of maize kernel breakage is crucial for quality control.

Purpose of the Study:

• To develop a quantitative detection model for maize kernel breakage rate using machine vision and deep learning.
• To create an improved Transformer-based deep learning model (MSA Transformer) for enhanced feature extraction and classification.
• To provide a reliable method for assessing food quality in agricultural products.

Main Methods:

• Extracted 27 geometric, shape, color, and texture features from maize kernel images.
• Developed an improved Transformer model (MSA Transformer) with multi-scale feature fusion and attention mechanisms.
• Utilized parallel branches for multi-granularity feature extraction and global/local attention for salient information enhancement.

Main Results:

• The MSA Transformer achieved 98.03% classification accuracy, surpassing the standard Transformer by 2%.
• Achieved high performance metrics: 99.13% average precision, 98.03% recall, and 97.87% F1-score.
• Quantitative detection model showed strong correlation (R²=0.9887) with actual measurements, with a low relative error (~6%).

Conclusions:

• The developed MSA Transformer model provides accurate and efficient quantitative detection of maize kernel breakage.
• Color and texture features are key for classification, while geometric features are important for mass prediction.
• This research establishes a foundation for real-time, quantitative food quality assessment in agricultural production.