Contrastive learning with prototype reliability calibration for plant leaf recognition

Plant leaf recognition is a fine-grained classification task in which visually similar species produce difficult negatives, while mislabeled, atypical, or stale memory entries that become prototype-inconsistent can receive excessive influence under similarity-only hard-negative mining. We propose ContrastPlant, a supervised contrastive training method with prototype-reliability calibration (PRC). A bounded class-partitioned memory supplies cross-batch candidates, an exponential moving average class anchor stabilizes the positive relation, and PRC weights each retrieved negative according to both its consistency with its labeled class and the ambiguity between its class and the query class. This preserves instance-level difficulty while reducing the contribution of prototype-inconsistent negatives. Experiments are conducted on Flavia, Swedish Leaf, Folio, and a fixed LeafSnap laboratory-to-field protocol. On Flavia, PRC obtains $$96.03\\pm 0.31$$ % accuracy, compared with $$93.31\\pm 0.16$$ % for cross-entropy, $$94.84\\pm 0.16$$ % for SupCon, and $$95.68\\pm 0.19$$ % for similarity-only Top-K with EMA. PRC also improves accuracy on Swedish Leaf and Folio, and achieves the largest practical gain under the more challenging LeafSnap field setting, improving fixed lab-to-field accuracy from $$71.49\\pm 0.62$$ % to $$75.87\\pm 0.67$$ %. Additional ablations, label-noise stress testing, representation-space statistics, modern-backbone experiments, and Jetson Orin Nano measurements support the effectiveness and deployment efficiency of the proposed training-time calibration.