Non-Destructive Detection of Chilled Mutton Freshness Using a Dual-Branch Hierarchical Spectral Feature-Aware Network

Area of Science:

• Food Science and Technology
• Spectroscopy
• Artificial Intelligence

Background:

• Accurate detection of meat freshness is crucial for food safety and quality control.
• Hyperspectral imaging offers rich spectral information for freshness assessment, but raw data requires sophisticated processing.
• Challenges include data redundancy, overlapping spectral features, and imbalanced sample distributions.

Purpose of the Study:

• To develop a high-accuracy method for detecting chilled mutton freshness using hyperspectral imaging and deep learning.
• To propose a novel Dual-Branch Hierarchical Spectral Feature-Aware Network (DBHSNet) to overcome limitations in existing methods.
• To enhance the precision of real-time quality monitoring in cold-chain meat systems.

Main Methods:

• Multi-stage data processing to enhance spectral purity and minimize redundancy.
• Development of DBHSNet incorporating a PBCA module for improved feature interaction between global and local branches.
• Integration of a task-driven MSMHA module for effective feature fusion and capturing spectral dynamics.
• Application of dynamic loss weighting to balance classification performance.

Main Results:

• DBHSNet achieved up to 7.59% higher accuracy in mutton freshness detection compared to conventional methods.
• The proposed network demonstrated superior weighted metrics, indicating robust performance.
• Enhanced awareness of sequential spectral bands and improved capture of fine-grained spectral details were observed.

Conclusions:

• DBHSNet provides a novel and effective approach for precise chilled mutton freshness detection.
• The study offers valuable support for real-time quality monitoring in cold-chain meat systems.
• This integration of hyperspectral imaging and advanced deep learning shows significant promise for food safety applications.