Proto-DS: A Self-Supervised Learning-Based Nondestructive Testing Approach for Food Adulteration with Imbalanced...

Area of Science:

• Food Science
• Machine Learning
• Spectroscopy

Background:

• Conventional food fraud detection using hyperspectral imaging (HSI) relies on machine learning but struggles with imbalanced datasets common in real-world applications.
• Imbalanced data leads to suboptimal performance, where minority classes are overshadowed by majority classes, posing a significant research challenge.
• Developing effective classifiers for small-scale, imbalanced datasets without dominant class bias is critical for reliable food fraud detection.

Purpose of the Study:

• To propose a novel, nondestructive detection approach, the Dice Loss Improved Self-Supervised Learning-Based Prototypical Network (Proto-DS), to address imbalanced learning challenges in HSI-based food fraud detection.
• To mitigate label bias from the most frequent class and enhance the robustness of the detection model.
• To validate the effectiveness of Proto-DS on diverse imbalanced hyperspectral food image datasets.

Main Methods:

• Developed the Proto-DS method, integrating self-supervised learning, prototypical networks, and Dice loss to handle imbalanced data.
• Collected and utilized three hyperspectral food image datasets with varying degrees of data imbalance: Citri Reticulatae Pericarpium (Chenpi), Chinese herbs, and coffee beans.
• Compared Proto-DS against state-of-the-art imbalanced learning techniques like Synthetic Minority Oversampling Technique (SMOTE) and class-importance reweighting, as well as conventional ML models (LMT, MLP, CNN).

Main Results:

• Proto-DS consistently outperformed conventional approaches and other imbalanced learning techniques across all tested datasets.
• Achieved an average balanced accuracy of 88.18%, significantly higher than Logistic Model Tree (59.42%), Multi-Layer Perceptron (60.38%), and Convolutional Neural Network (66.34%).
• Demonstrated that self-supervised learning is key to improving imbalanced learning, with prototypical networks and Dice loss further enhancing classification performance.

Conclusions:

• The proposed Proto-DS method offers a superior solution for nondestructive food fraud detection, particularly effective on imbalanced datasets.
• Self-supervised learning provides complementary information and is crucial for enhancing imbalanced learning performance in HSI applications.
• Combining self-supervised learning with prototypical networks and Dice loss presents a promising strategy for building robust models with limited training data.