GAGE: A Graph-based and Prototype-Guided Framework for Fine-Grained Visual Ginseng Age Estimation

• Builds a real-world benchmark for fine-grained ginseng age estimation under practical constraints. • Develops a graph-based visual representation to capture complex biological morphology. • Introduces a prototype-guided learning strategy to improve regression robustness. • Achieves state-of-the-art accuracy on challenging ginseng age estimation benchmarks. Accurate age estimation of ginseng is critical for market pricing and quality grading, yet remains challenging due to the non-Euclidean morphology and high intra-class variability of biological samples. Manual assessment is subjective, while physicochemical detection is destructive and costly. Existing computer vision approaches, largely developed for structured objects such as human faces, often struggle to capture the irregular topological dependencies inherent in ginseng, including rhizome scar stacking and stochastic root branching patterns. To address these challenges, we propose GAGE, a graph-based and prototype-guided framework for fine-grained visual ginseng age estimation. First, we construct a dataset with expert-validated continuous age annotations covering the 15–30 year growth range. Second, we introduce a Graph Attention Visual Encoder (GAVE) that reorganizes local visual patches into a feature-driven graph, enabling explicit modeling of long-range structural relationships beyond regular grid representations. Third, to mitigate regression instability under limited and imbalanced data, we propose a retrieve-and-augment prototype learning paradigm, where a hierarchically constructed prototype library provides contextual references for fine-grained prediction. Extensive experiments on the full dataset demonstrate that GAGE achieves competitive performance, with a Mean Absolute Error (MAE) of 0.98 ± 0.02 years and a tolerance accuracy (CS@2) of 92.65% ± 0.74. Compared with representative methods, the proposed framework shows consistent improvements, particularly in constraining predictions within practical error bounds. These results suggest that GAGE provides an effective and non-destructive approach for automated ginseng age estimation in real-world scenarios.