The automated detection of persimmons in orchards is hindered by two primary factors: fruits of the same cultivar possess nearly identical color, causing adjacent instances to be easily merged, and heavy occlusion by leaves and stems hides large portions of the fruit surface. These challenges often co-occur and severely degrade detection performance. To address them jointly, we propose PersimmonDet, a collaborative framework that integrates a dual-path modulation fusion module (MFM) for color–occlusion decoupling, a modulation-guided dynamic upsampling operator (M-DySample) that preserves boundaries under occlusion, a scale-aware lightweight detection head (MB_Head), and a repulsion-enhanced Wise-IoU loss. The key innovation is an implicit coordination mechanism: the loss function generates strong gradients based on difficult samples, and these gradients flow back to train the fusion and upsampling modules, forming a closed optimization loop. Experiments on a self-built persimmon dataset show that PersimmonDet achieves an mAP@0.5 of 96.13% and mAP@0.5:0.95 of 84.42%, outperforming multiple lightweight detectors. The largest gains occur based on heavily occluded and densely clustered fruits. Ablation studies reveal a clear improvement when all four components are combined, confirming their complementary contributions. The same architecture, retrained on a winter jujube dataset without modification, delivers consistent advantages, demonstrating cross-crop transferability. These results indicate that combining problem-aware modulation, boundary-preserving upsampling, and difficulty-focused training is a promising approach for robust fruit detection in orchard environments.

In complex orchard environments, sweet persimmons tend to grow in dense clusters and display similar coloration across different maturity stages, leading to heavy occlusion and poor inter-class color discriminability. To address these challenges, this paper presents RCPNet, a detection network tailored for such field conditions. The model integrates a Rectangular Self-Calibration Module (RCM) and a Context Feature Calibration Gating (CFCG) module. RCM strengthens axial context capture, while CFCG improves feature calibration; together they reduce local feature ambiguity and help reconstruct missing information in occluded regions. For distinguishing fruits at different ripening stages that share similar colors, a Parallelized Patch-aware Attention (PPA) detection head is adopted. By leveraging self-attention and multi-branch strategies, this head suppresses feature degradation and notably enhances sensitivity to color contrast. Experiments on sweet persimmon images show that RCPNet improves mean Average Precision (mAP) by 2.7 percentage points and mAP@0.5:0.95 by 3.7 percentage points over the baseline, reaching 93.6% detection accuracy for immature fruits. Ablation studies and comparisons with mainstream detectors indicate that the proposed model, though slightly heavier than lightweight detectors of analogous capacity, surpasses the accuracy of a larger small-scale counterpart and exhibits satisfactory robustness. Strong performance on a self-collected flat jujube dataset further confirms its generalization ability. The method delivers highly accurate detection for occluded and near-color fruits, providing technical support for precise fruit recognition and automated picking.