Proto-DS: Advanced AI for Food Testing and Quality Control

By Muhammad OsamaReviewed by Susha Cheriyedath, M.Sc.Nov 19 2024

In a recent study published in the journal Foods, researchers introduced the Dice Loss Improved Self-Supervised Learning-Based Prototypical Network (Proto-DS), a novel framework designed to detect food adulteration using hyperspectral imaging (HSI).

Advancements in HSI Technology

HSI is a powerful analytical technique that captures a wide range of light spectra for each pixel in an image, enabling detailed spectral analysis of materials. Its non-destructive, rapid, and precise assessment capabilities make it valuable for fields like agriculture, food science, and environmental monitoring.

In food testing, HSI is particularly effective for detecting adulteration and verifying the authenticity of food products by distinguishing between genuine and adulterated items based on their unique spectral signatures.

However, conventional machine learning methods applied to HSI often assume balanced class distributions, which are rarely found in real-world applications. This imbalance can significantly reduce detection performance, particularly for less common classes that are overshadowed by more frequent ones. To address this issue, the study introduced Proto-DS, a framework aimed at enhancing detection accuracy for imbalanced datasets, thereby improving the reliability and effectiveness of HSI in practical scenarios.

Proto-DS: A Novel Approach for Imbalanced HSI Data

In this paper, the authors aimed to develop an effective classifier for detecting food adulteration in hyperspectral images, especially in cases with highly imbalanced datasets. They proposed Proto-DS, a novel method combining self-supervised learning, Dice loss, and prototypical networks to reduce the label bias caused by dominant classes. This innovative approach aims to enhance classifier accuracy and reliability with limited training data.

To validate the effectiveness of Proto-DS, three hyperspectral food image datasets were used: Chinese herbs, Citri Reticulatae Pericarpium (chenpi), and coffee beans. These datasets had different levels of class imbalance, providing a robust testing environment for the method. The researchers also compared Proto-DS with leading techniques like class-importance reweighting and the Synthetic Minority Oversampling Technique (SMOTE).

The training methodology consisted of a two-step process. Initially, the model applied spectral contrastive learning to extract meaningful features from the image without supervision. Then, it was fine-tuned on the labeled imbalanced data using the Dice loss function, which is known for effectively handling class imbalance in classification tasks.

Key Findings and Insights

The study showed that Proto-DS significantly outperformed traditional models in terms of balanced accuracy, achieving an impressive average accuracy of 88.18 % across various training sample sizes. In contrast, conventional models such as the logistic model tree (LMT), convolutional neural network (CNN), and multi-layer perceptron (MLP) achieved balanced accuracies of only 59.42 %, 66.34 %, and 60.38 %, respectively. These outcomes highlighted the effectiveness of Proto-DS in handling imbalanced datasets.

The experiments highlighted the crucial role of self-supervised learning in enhancing classifier performance on imbalanced datasets. By leveraging complementary information, Proto-DS effectively reduced label bias and better represented minority classes. Combining self-supervised learning with Dice loss further improved classification accuracy while minimizing overfitting to the majority class.

Additionally, the authors emphasized the importance of utilizing all effective near-infrared (NIR) spectra from hyperspectral images instead of relying on averaged spectra. This approach preserved data variability, allowing the model to train on a larger and more diverse dataset, which in turn improved generalization and detection performance in food testing.

This research provides valuable insights into food safety and quality control. Proto-DS offers an adaptable approach for detecting food adulteration across various commodities, enabling food manufacturers and regulators to verify product authenticity and protect consumer health. The integration of self-supervised learning with advanced machine learning techniques also presents opportunities for developing systems capable of performing well with limited training data.

Conclusion and Future Directions

In summary, Proto-DS addresses the challenges of imbalanced datasets in HSI-based food testing. By combining self-supervised learning, Dice loss, and prototypical networks, it sets a new benchmark for non-destructive food testing methods.

Future work should focus on refining Proto-DS and exploring its potential in pharmaceuticals and environmental monitoring. Expanding the validation datasets and incorporating soft labels for regression tasks could also enhance its adaptability and effectiveness, ultimately improving food testing standards worldwide.

Journal Reference

Pang, K.; Liu, Y.; Zhou, S.; Liao, Y.; Yin, Z.; Zhao, L.; Chen, H. Proto-DS: A Self-Supervised Learning-Based Nondestructive Testing Approach for Food Adulteration with Imbalanced Hyperspectral Data. Foods 2024, 13, 3598. DOI: 10.3390/foods13223598, https://www.mdpi.com/2304-8158/13/22/3598

Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.