Related Experiment Video
Updated: Jan 12, 2026

Integration of Animal Behavioral Assessment and Convolutional Neural Network to Study Wasabi-Alcohol Taste-Smell Interaction
Published on: August 16, 2024
Evaluating the pungency perception of Zanthoxylum armatum DC. oil by facial recognition techniques and machine
Shijun Zheng1, Qixin Kan1, Guo Liu1,2
1Guangdong Provincial Key Laboratory of Nutraceuticals and Functional Foods, College of Food Science, South China Agricultural University, Guangzhou, 510642, China.
Abstract:
Zanthoxylum armatum DC. (ZAD) oil is a kind of popular seasoning for its unique flavor and numbing taste. Pungency is characterized as a numbing and stimulating sensation in ZAD oil. This study aimed to objectively assess ZAD oil pungency by combining sensory evaluations with facial expression analysis. The contents of four major pungency compounds, including hydroxy-α-sanshool, hydroxy-β-sanshool, hydroxy-ε-sanshool, and hydroxy-γ-sanshool, were determined. Among them, the content of hydroxy -α-sanshool was 242.741 ± 18.293 mg/mL. The pungency intensity increased with higher Scoville pungency units (SPU). There was a demarcation point for the degree of preference. That is, before Sample 2 (with an SPU of 11,200), the degree of preference showed an upward trend as the numbing degree increased. After Sample 2 (with an SPU of 11,200), the degree of preference decreased as the numbing degree increased. Moreover, facial expression recognition technology was employed to collect data on 20 facial action units (AUs) from 52 participants. Subsequent correlation analysis highlighted significant AUs associated with the sensation of pungency, including AU4 (brow lowering), AU6 (cheek raising), AU7 (lid tightening), AU10 (upper lip raising), AU23 (lip tightening), AU25 (lips parting), and AU26 (jaw dropping). Furthermore, utilizing the Extra Trees model, the study achieved 75.5% accuracy in ternary and 90.4% in binary classification of pungency. SHapley Additive exPlanations methodology was applied to clarify feature contributions, enhancing model transparency. This research showed the potential of facial recognition techniques to evaluate pungent flavors of ZAD oil.

