Abstract:

To address the complex scenarios of apple fruit detection in orchard environments， such as target overlapping， uneven illumination and varying scales， and to meet the balance requirement between detection accuracy and computational resources for model deployment， a lightweight apple fruit detection model YOLO-AP based on the improvement of YOLO11n was proposed. Firstly， by integrating GhostModule and dynamic convolution to improve the feature extraction module， a GD_C3K2 module was proposed to enhance the model’s feature extraction capability while reducing model complexity. A global-local dual-stream feature fusion network was designed to reconstruct the neck network by simplifying the bidirectional feature pyramid network and combining the adaptive-downsamping module. Additionally， a global-to-local spatial aggregation module was introduced in the network head to further enhance the model’s global and local spatial modeling capabilities. The lightweight shared convolution head RepCoord-LDH was constructed by combining the coordinate attention mechanism and reparameterized convolution to reduce model complexity while maintaining high detection accuracy. Finally， WiseIOUv3 was introduced as the model’s bounding box loss function to optimize the bounding box regression performance. The results showed that YOLO-AP achieved a precision of 89.1%， a recall of 88.9% and a mean average precision of 96.1%， which were 0.5， 0.1 and 1.1 percentage points improvements compared to the baseline model YOLO11n. Meanwhile， the model’s number of parameters and floating-point operations are 1.4 M and 3.6 G， respectively， which were only 53.8% and 54.5% of the baseline model. In the comparison with mainstream detection algorithms， YOLO-AP also outperformed other algorithms in terms of detection performance and model complexity. In conclusion， the proposed YOLO-AP model provided an effective technical reference for apple detection in complex environments.

Key words: apple fruit recognition, YOLO11, lightweight, target detection, deep learning

摘要：

针对果园环境下苹果果实目标重叠、光照不均且尺度不一等复杂场景特点，同时为满足模型落地部署时检测精度与计算资源之间的平衡需求，提出一种基于YOLO11n改进的轻量化苹果果实检测模型YOLO-AP。首先，结合幽灵卷积与动态卷积改进特征提取模块，提出一种GD_C3K2模块，增强模型特征提取能力的同时降低模型复杂度；设计一种全局-局部双流特征融合网络，通过简化双向特征金字塔网络与自适应下采样模块对颈部网络进行重构，并在网络首部上引入全局到局部空间聚合模块进一步增强模型全局和局部空间建模能力；结合坐标注意力机制和重参数化卷积构建轻量化共享卷积头RepCoord-LDH，减少模型复杂度的同时维持高检测精度；最后，引入WiseIOUv3作为模型边界框损失函数，优化边界框回归性能。结果表明，YOLO-AP的精确率、召回率和平均精确率分别达到89.1%、88.9%和96.1%，相较于基线模型YOLO11n分别提升0.5、0.1和1.0百分点，同时模型的参数量和浮点运算量分别为1.4 M和3.6 G，仅为基线模型的53.8%和54.5%。与主流检测算法对比，YOLO-AP在模型检测性能和复杂度等指标中也优于其他算法。综上所述，YOLO-AP模型可为复杂环境下的苹果检测提供有效的技术支持。

关键词: 苹果果实识别, YOLO11, 轻量化, 目标检测, 深度学习