Hyperspectral Imaging and Grading of Kiwifruit with Hierarchical 3D Convolution Data Processing.

Area of Science:

• Agricultural Science
• Computer Vision
• Food Science

Background:

• Kiwifruit quality, particularly sugar content, significantly impacts consumer satisfaction and purchase decisions.
• Current destructive sampling methods for kiwifruit grading are inefficient and unsuitable for large-scale production.
• Accurate, rapid grading is essential for maintaining product quality and market competitiveness.

Purpose of the Study:

• To develop a non-destructive kiwifruit classification method based on sugar content.
• To leverage hyperspectral imaging and deep learning for precise quality assessment.
• To establish an efficient grading system for modern agricultural practices.

Main Methods:

• Proposed a novel Hierarchical 3D Convolution and Attention Mechanism Network (H3DAMNet) for kiwifruit classification.
• Utilized hyperspectral data, applying 3D convolution for deep spatial-spectral feature extraction.
• Incorporated channel attention and bottleneck self-attention mechanisms to enhance feature relevance and global information modeling.

Main Results:

• The H3DAMNet achieved an overall accuracy (OA) of 97.5% and an average accuracy (AA) of 97.3% on a test set of 280 kiwifruit.
• Successfully classified kiwifruit into three grades based on industry-standard sugar content levels (≥14.5 °Brix, 13.5-14.5 °Brix, ≤13.5 °Brix).
• Demonstrated the effectiveness of the proposed deep learning model for non-destructive fruit quality assessment.

Conclusions:

• The H3DAMNet offers an accurate and efficient non-destructive method for kiwifruit grading based on sugar content.
• This approach provides a valuable reference for the classification of similar fruits using hyperspectral imaging and deep learning.
• The study highlights the potential of advanced AI techniques in revolutionizing agricultural quality control.