Comparative Assessment of Hyperspectral Image Segmentation Algorithms for Fruit Defect Detection Under Different...

Area of Science:

• Agricultural Engineering
• Computer Vision
• Spectroscopy

Background:

• Fruit quality assessment is crucial for supply chain management.
• Non-contact, high-throughput inspection systems are needed to reduce postharvest losses.
• Hyperspectral imaging offers rich spectral information for defect detection.

Purpose of the Study:

• To comparatively analyze hyperspectral image segmentation algorithms for fruit defect detection.
• To evaluate algorithm performance under diverse illumination conditions.
• To identify optimal acquisition protocols for improved segmentation accuracy.

Main Methods:

• Evaluated four segmentation algorithms: Spectral Angle Mapper, Random Forest, Support Vector Machine, and Neural Network.
• Tested algorithms under local, simultaneous, and sequential illumination modes.
• Acquired hyperspectral data from tomato fruit samples (450-850 nm).
• Assessed performance using metrics like accuracy, precision, recall, F1-score, and IoU.

Main Results:

• Random Forest demonstrated superior performance across most segmentation metrics.
• Highest accuracy (0.9971) achieved by Random Forest under sequential illumination.
• Best F1-score (0.8996) and IoU (0.8176) obtained under simultaneous illumination.
• Neural Network showed competitive results; Spectral Angle Mapper was sensitive to illumination but memory-efficient.

Conclusions:

• Acquisition protocol optimization significantly enhances segmentation performance for fruit defect detection.
• Results support the development of accurate, non-contact, high-throughput inspection systems.
• Improved quality control can reduce postharvest losses and enhance supply chain efficiency.