LUCID: A Training-Free Few-Shot Framework for Leaf Disease Diagnosis via Visual–Knowledge Evidence Fusion

The diagnosis of leaf diseases in medicinal and non-mainstream plants is often hindered by the scarcity of labeled samples, whereas existing deep learning methods generally rely on large-scale labeled datasets. Prototype-based few-shot methods are constrained by the limited number of support samples and fine-grained visual similarities among categories, which can result in unstable class representations. To address these limitations, LUCID, a training-free few-shot framework, is proposed for diagnosing leaf diseases in medicinal and non-mainstream plants. The framework constructs category-level visual prototypes from a small number of support images and integrates them with structured symptom knowledge to establish complementary discriminative criteria. LUCID employs a vision-dominated confidence-gated fusion mechanism that adjusts the knowledge correction weights according to the relative discriminative confidence of the visual and knowledge branches. For candidate categories that remain difficult to distinguish after the visual and knowledge scores have been fused, Top-2 difference-driven re-examination is further performed to incorporate additional discriminative information concerning key symptom differences. The prediction ranking is then revised based on the newly incorporated information, and a concise diagnostic rationale is ultimately generated. Few-shot experiments were conducted on the Azadirachta indica leaf dataset from the publicly available AI-MedLeafX dataset, and LUCID was compared with supervised vision models, vision-language models, and other methods. The experimental results showed that LUCID achieved the best performance across all few-shot settings. In the 1-shot setting, its Accuracy and Macro-F1 exceeded those of the best-performing baseline by 15.68 and 14.22 percentage points, respectively. Beyond the primary dataset, LUCID also demonstrated consistent performance advantages on the Ocimum tenuiflorum and Ziziphus jujuba datasets, validating its ability to generalize across plant species. These results indicate that LUCID can effectively integrate limited visual evidence with structured symptom knowledge without additional training, thereby providing an accurate, generalizable, and evidence-based solution for diagnosing leaf diseases in medicinal and non-mainstream plants.