| 陈炜炜,郭政,严翔,唐碧瑜,郑永强.基于YOLOv8n的典型场景下柑橘花检测算法 [J].中国南方果树,2026,55(3): |
| 基于YOLOv8n的典型场景下柑橘花检测算法 |
| Citrus Flower Detection Algorithm Based on YOLOv8n in Typical Scenarios |
| 投稿时间：2025-02-28 修订日期：2025-03-17 |
| DOI：10.13938/j.issn.1007-1431.20250100 |
| 中文关键词: 柑橘花量 YOLOv8n 目标检测 轻量化 注意力机制 |
| 英文关键词:Citrus flowers quantity YOLOv8n Target detection Lightweight Attention mechnism |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| 摘要点击次数: 133 |
| 全文下载次数: 0 |
| 中文摘要: |
| 【目的】柑橘花量的精准管理是确保果园产量和品质的重要环节，其关键在于快速精确的花朵检测技术。因此，本研究提出了一种基于YOLOv8n改进的YOLO-CF柑橘花识别网络。【方法】 针对原始网络模型做了如下改进：首先，引入FasterNet骨干网络替换原始模型的骨干网络，降低初始模型的复杂度，减少了原有模型冗余运算量，并使改进后模型体积进一步缩小，便于后续在边缘设备进行部署实现农业场景移动实时监测。其次，在骨干网络和特征融合层嵌入CA（Coordinate Attention）空间注意力模块，通过引入双向池化机制，分别在水平和垂直方向分解空间坐标，构建跨通道注意力权重图，实现特征图长程依赖关系的建模，有效增强模型对柑橘花形态特征的判别能力，同时在特征融合层建立位置信息增强通道，显著提升了密集遮挡、光照不均及多尺度背景干扰等复杂农业场景下的目标定位精度。【结果】试验结果表明：与Fater R-CNN和原始YOLOv8n网络模型相比，YOLO-CF网络模型平均精度分别提升4.7%和1.9%，召回率分别提升了2.3%和1.3%，且大小分别减少95.4%和19.0%。经过改进后的网络模型在执行遮挡、密集和小目标检测任务时平均精度分别提升1.9%、3.6%和2.0%，同时未出现漏检和误检情况且检测速度提升33.7%。【结论】提出的YOLO-CF网络模型能够在复杂自然环境下对柑橘花进行精准检测，并且具有较快的运算效率，可以为柑橘花量智能监测技术提供支撑。 |
| 英文摘要: |
| 【Objective】The meticulous management of citrus flower volume constitutes a critical step in optimizing the yield and quality of orchards. Therefore, a YOLO-CF citrus blossom recognition network based on YOLOv8n improvement is proposed in this study. 【Methods】TThe following improvements are made to the original network model: first, the FasterNet backbone network is introduced to replace the backbone network of the original model, which reduces the complexity of the initial model, reduces the amount of redundant operations of the original model, and further reduces the size of the improved model, which is easy to be deployed in the edge devices to realize the real-time monitoring of the agricultural scene on the move.Second, the CA (Coordinate Attention) spatial attention module is embedded in the backbone network and the feature fusion layer, and by introducing a two-way pooling mechanism, the spatial coordinates are decomposed in the horizontal and vertical directions respectively, and a cross-channel attention weight map is constructed to realize the modeling of long-range dependency relationship of the feature map, which effectively enhances the model's ability to dis-criminate the morphological features of citrus flowers, and meanwhile, the location information enhancement channel is established in the feature fusion layer to significantly improve the target positioning accuracy in complex agricultural scenes such as dense occlusion and multi-scale background interference. At the same time, the position information enhancement channel is established in the feature fusion layer, which significantly im-proves the target localization accuracy in complex agricultural scenarios such as dense occlusion, uneven illu-mination and multi-scale background interference.【Results】 The experimental results show that compared with the Fater R-CNN and the original YOLOv8n network model, the YOLO-CF network model improves the average precision by 4.7% and 1.9%, the recall by 2.3% and 1.3%, and the size reduction by 95.4% and 19.0%, respec-tively. The improved network model improves the average precision by 1.9%, 3.6%, and 2.0% for occlusion, dense, and small target detection tasks, respectively, with no leakage and no false detection and a 33.7% increase in detection speed.【Conclusions】 The proposed YOLO-CF network model can accurately detect citrus blossoms in complex natural environments and has a fast computing efficiency, which can support the intelligent moni-toring technology of citrus blossom quantity. |
| 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
