Predicting cookie visual appetite with interpretable machine learning

Visual appearance shapes consumers' pre-ingestive evaluation of baked foods, yet machine-vision studies have rarely linked interpretable appearance descriptors directly with consumer visual appetite. This study developed an interpretable machine-vision framework for commercial handmade cookies. A total of 159 standardized images from 51 cookie categories were evaluated by 700 adults, yielding 6829 valid visual-appetite responses. Sixty descriptors spanning color, texture, morphology, spatial color distribution, and surface characteristics were extracted. A 12-feature RBF-SVM achieved an Accuracy of 0.918 and a Macro F1 of 0.915 under sample-level nested cross-validation. SHAP analysis showed that greater local texture complexity and redder hue were associated with stronger contributions toward the High-appetite grade, whereas higher image-derived Browning Index values were generally associated with lower contributions. Glossiness and Homogeneity exhibited preferred-range relationships. Emotion-dimension and CATA responses showed consistent perceptual patterns, while category-disjoint validation indicated that transfer to unseen cookie categories was more difficult. These results provide interpretable, measurable appearance variables for cookie appearance evaluation and product development.