Pre-trained Vision-Language Models for smartphone-based assessment of cucumber freshness: validation against sensory evaluation

Scritto il 15/11/2026
da Sina Keyhani

Curr Res Food Sci. 2026 Nov 4;13:101617. doi: 10.1016/j.crfs.2026.101617. eCollection 2026.

ABSTRACT

This study investigated whether publicly accessible pre-trained Vision-Language Models (VLMs), GPT-5.5 (OpenAI) and Gemini 3.1 Pro (Google), can accurately assess cucumber (Cucumis sativus L.) freshness directly from conventional smartphone images acquired under practical imaging conditions. Cucumbers were stored at 4 °C for 21 days and evaluated on days 0, 4, 11, 17, and 21. Images were captured under controlled, indoor ambient, and outdoor natural lighting, with and without commercial cellophane wrapping. AI-generated freshness scores were validated against a human sensory reference and physicochemical, mechanical, and visual quality attributes, including firmness, weight loss, peduncle thickness, and CIELAB color parameters, together with visible changes in surface condition. Firmness and peduncle thickness decreased significantly during storage, whereas weight loss, yellowness, and other visible signs of quality deterioration increased. Hierarchical cluster and Pearson correlation analyses demonstrated strong associations between VLM predictions, storage progression, and objective quality indicators. Importantly, no statistically significant differences in freshness predictions were detected across the tested lighting conditions and packaging treatments (p > 0.05), although the relatively small sample size limits the interpretation of these findings. External validation and repeated-query analyses further supported consistent model performance across diverse imaging environments. These findings demonstrate that pretrained VLMs can reliably evaluate cucumber freshness from ordinary smartphone images without specialized sensors, handcrafted feature extraction, or model retraining, highlighting their potential as scalable, low-cost tools for postharvest quality monitoring, retail inspection, quality-based dynamic pricing during storage, and food waste reduction throughout fresh produce supply chains.

PMID:43034735 | PMC:PMC13769528 | DOI:10.1016/j.crfs.2026.101617