An image-based food quality analysis framework driven by large language models.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Food Science

Background:

• Traditional machine learning (ML) and deep learning (DL) for food quality analysis require extensive data and expertise.
• Large language models (LLMs) present new possibilities for multimodal perception and reasoning.

Purpose of the Study:

• To develop an LLM-driven vision inspection framework for food quality analysis.
• To evaluate the performance of LLM approaches against traditional methods.

Main Methods:

• Developed a framework with two inference routes: text-only inference (TOI) and vision-language inference (VLI).
• Integrated a graphical user interface (GUI) for automated feature extraction and data generation.
• Evaluated state-of-the-art LLMs on food ripeness, freshness, and authenticity datasets.

Main Results:

• The VLI approach achieved high accuracy (0.964-1.000) with fewer training images.
• The TOI approach attained 1.000 accuracy on simple tasks and 0.875 on complex ones.
• LLM approaches matched or surpassed traditional ML/DL baselines.

Conclusions:

• LLM-driven vision inspection is feasible and practical for food quality analysis.
• The proposed framework enables efficient, fine-tuning-free food quality assessment.
• Prompt-based in-context learning eliminates the need for task-specific fine-tuning.