基于改进YOLO 11s的轻量化百香果检测模型

Lightweight Passion Fruit Detection Model Based on Improved YOLO 11s

摘要:

为满足移动端设备在自然果园环境下对百香果实时检测的轻量化需求,本文提出一种基于改进 YOLO 11s 的轻量化检测模型———MEC-YOLO 11s。 该模型利用 MobileNetV4 轻量化网络替换骨干网络(Backbone),旨在大幅压缩模型参数规模的同时维持对多尺度特征的高效捕获;在此基础上嵌入 ECA 注意力机制,通过局部跨通道交互增强模型对果实纹理与颜色特征的响应,有效抑制复杂背景下的噪声干扰;结合设计的 CSPStage 模块优化颈部特征融合网络(Neck),利用梯度分流策略提升特征聚合效率以解决重叠遮挡问题。 试验结果表明,MEC-YOLO 11s 模型在自制数据集上的平均精度均值(mAP)达到 94. 17% ,检测速度高达 132. 0 f / s;与基线模型相比,在保持检测精度相当的前提下,参数量和浮点运算量分别降低了 49. 1% 和 58. 1% ,能够满足资源受限移动端设备的实时部署需求。

Abstract:

In order to meet the requirements for lightweight and real-time passion fruit detection on mobile devices in natural orchard environments, a lightweight detection model based on the improved YOLO 11s architecture, designated as MEC-YOLO 11s, was proposed. Firstly, the lightweight MobileNetV4 was introduced to replace the original backbone network, which significantly reduced the model parameter scale while maintaining efficient feature capture capabilities across multiple scales. Simultaneously, the efficient channel attention (ECA )mechanism was embedded into the feature extraction network. By facilitating local cross-channel interaction, this mechanism enhanced the model??s sensitivity to fruit texture and color, effectively suppressing noise interference from complex backgrounds. Additionally, a CSPStage module was implemented to optimize the neck feature fusion network, utilizing a gradient shunting strategy to improve feature aggregation efficiency and address occlusion problems. Experimental results demonstrated that the MEC-YOLO 11s model achieved significant improvements in comprehensive performance. The model achieved a mean average precision (mAP)of 94. 17% with a detection speed of 132. 0 f / s on the self-constructed dataset. Compared with the baseline YOLO 11s model, the parameter count and floating-point operations (FLOPs)were reduced by 49. 1% and 58. 1% , respectively, while maintaining comparable detection accuracy. These findings validated that the proposed model can meet the real-time deployment requirements of resource-constrained mobile terminals, providing a technical foundation for future intelligent harvesting systems.