Unified Multi-Level Grape Counting Framework with Low Annotation Cost for Natural Orchard Scenes

Accurate multi-level grape counting in natural orchard images provides useful visual information for image-based crop-load characterization and vineyard management. However, cluster overlap, occlusion, boundary adhesion, and illumination variation complicate berry localization and cluster-wise counting. Existing detection- and density-regression-based methods remain susceptible to missed detections, false positives, and cross-cluster misassignment, whereas approaches relying on pixel- or instance-level supervision typically require costly annotations. To address these challenges, this paper proposes UniGC, a unified framework for multi-level grape counting. At the berry level, UniGC employs CSD-P2PNet, a P2PNet-based point-regression branch equipped with a Condition-Structure Decoupled Frequency Enhancement Block (CSD-Block), to improve the localization and counting of image-visible berries under complex illumination and occlusion using only berry point annotations. At the cluster level, a You Only Look Once (YOLO) detector predicts grape-cluster bounding boxes, which are used as prompts for the Segment Anything Model (SAM) to generate cluster instance masks. We further propose Boundary-Depth Guided Mask-Point Assignment (BDMPA), which assigns each predicted berry point to at most one grape-cluster instance and thereby produces cluster-wise berry counts. Selected key experiments on the mixed dataset organized from WGISD and GBISC were repeated using three random seeds. CSD-P2PNet achieved an MAE of 23.4407 ± 0.1577, an RMSE of 32.8898 ± 1.2682, and a point-localization F1 score of 0.8374 ± 0.0097. YOLO12-S achieved an F1 score of 0.8433 ± 0.0056 and an AP50 of 0.8789 ± 0.0080 for grape-cluster detection. For BpGC counting, UniGC obtained a matched-cluster MRD of 0.1288 ± 0.0043, a 1−FVU of 0.8589 ± 0.0098, and an all-cluster penalized mean relative deviation (ACP-MRD) of 0.5062 ± 0.0124. These results demonstrate that UniGC achieves competitive multi-level grape counting performance while requiring only point and bounding-box annotations for training.