Probabilistic Objectness Modeling With Common Features for Tea Bud Detection Across the Entire Harvest Cycle

ABSTRACT Throughout the entire tea bud harvest cycle, the buds undergo a continuous transition from emergence to maturity, accompanied by substantial morphological changes. These dynamic changes introduce significant intraclass variations, under which traditional tea bud detection algorithms often exhibit limited generalization capability and reduced stability. To reduce false negatives caused by morphological variations across different maturation stages, this study builds upon LW‐DETR and proposes a probabilistic framework for tea bud objectness estimation. The framework improves confidence prediction by modeling common features shared among tea buds. First, a common feature encoder is introduced to project each embedding into a tea bud common feature subspace, where the representations are encouraged to learn shared characteristics and reduce interembedding feature discrepancies. Then, probability distribution estimation and objectness likelihood maximization for known tea bud objects are alternated within the common feature subspace. This iterative process enables the estimation of objectness probabilities for different bonding boxes, which are subsequently used to refine the original model's output. Experimental results show that, on the 25_ZC108 dataset split by collection date, the proposed method achieved the best performance across all 7 days, with a maximum AP50 of 85.10%. Compared to state‐of‐the‐art detection methods, the introduced probabilistic framework demonstrates superior detection accuracy and enhanced generalization capability, particularly on datasets characterized by substantial intra‐class variations. Notably, the highest recall improved by 31.13%, effectively reducing false negatives. Furthermore, the proposed approach achieved the best results on datasets collected across different years and tea cultivars, which attests to the robustness and effectiveness of the common‐feature‐based probabilistic objectness framework.