Identifying moldy peanut using hyperspectral imaging by correction of noisy labelling

Area of Science:

• Food Science
• Agricultural Engineering
• Computer Vision

Background:

• Aflatoxins in peanuts are a major human health concern.
• Hyperspectral imaging (HSI) and machine learning offer non-destructive detection of moldy peanuts.
• Noisy labels are a common challenge in HSI data labeling for moldy regions.

Purpose of the Study:

• To develop a label quality quantification framework (CL-ST) for identifying noisy labels in moldy peanut HSI.
• To improve the accuracy of aflatoxin detection in peanuts by addressing label noise.
• To assess the impact of noisy labels on model performance and feature selection.

Main Methods:

• Proposed a label quality quantification framework integrating confidence level and statistical testing (CL-ST).
• Estimated initial labeling confidence using out-of-sample probability.
• Assessed and ranked label quality via statistical testing, determining optimal noisy rate (NR) for model rebuilding.

Main Results:

• CL-ST reliably quantifies label quality and identifies noisy pixels in moldy peanut HSI.
• Rebuilt models using CL-ST achieved significantly improved accuracy (98.96%) and precision (97.09%).
• Noisy labels were identified as the primary cause of incorrect decision boundaries in models.

Conclusions:

• CL-ST effectively addresses noisy labels in HSI for moldy peanut detection.
• The framework enhances food quality assessment accuracy and reliability.
• CL-ST is a versatile, hyperparameter-free tool with broad applications in HSI data analysis.