YOLO-Apple：一种用于苹果幼果检测的改进型目标检测方法

YOLO-Apple: A Modified Object Detection Method for Apple Young-Fruit Detection

摘要: 针对自然果园场景中，苹果幼果因通体青绿色、与叶片色调相近，且本身尺度偏小、分布密集又易被枝叶遮挡，检测时易出现漏检与误检的问题，文章提出了一种面向苹果幼果检测的改进型目标检测模型YOLO-Apple。该方法以YOLOv11n为基线，在主干C2PSA模块引入轻量特征增强单元Mona (C2PSA_Mona)以强化复杂背景下的特征映射能力，改造C3k2模块并在其增强块嵌入EMA注意力(C3k2_EMA)，提升特征表达与选择能力。实验采用来自新疆阿克苏市与石河子市等自然场景采集的5558张苹果幼果图像数据集，按7:2:1划分为训练、验证与测试集。结果表明，YOLO-Apple在测试集上实现P = 91.5%、R = 75.9%、F1 = 83.0、mAP@0.5 = 84.4%的检测性能，在6.5 GFLOPs计算量与2.6 M参数量的轻量级开销下优于多款主流对比模型；消融实验验证了C2PSA_Mona与C3k2_EMA模块的改进有效性，为自然果园苹果幼果检测提供了新方法。

Abstract: To address the missed and false detection of young apple fruits in natural orchard scenes caused by their turquoise color, highly similar to leaf tones, along with small target scale, dense spatial distribution, and easy occlusion by branches and leaves, this paper proposes an improved object detection model, YOLO-Apple, for young apple fruit detection. Based on the YOLOv11n baseline, this method introduces the lightweight feature enhancement unit Mona into the backbone C2PSA module (denoted as C2PSA_Mona) to strengthen the feature mapping capability in complex backgrounds, and modifies the C3k2 module by embedding the EMA attention mechanism into its enhancement block (denoted as C3k2_EMA) to improve the feature representation and selection capabilities. Experiments were performed on a dataset of 5558 young apple fruit images collected from natural scenes in Aksu City and Shihezi City of Xinjiang, which was partitioned into training, validation and test sets at a ratio of 7:2:1. Experimental results show that YOLO-Apple achieves the detection performance of P = 91.5%, R = 75.9%, F1 = 83.0 and mAP@0.5 = 84.4% on the test set, and outperforms multiple mainstream comparison models with a lightweight overhead of 6.5 GFLOPs and approximately 2.6 M parameters. Ablation experiments further verify the effectiveness of the C2PSA_Mona and C3k2_EMA modules, thus providing a novel method for young apple fruit detection in natural orchard scenes.