基于改进 YOLOv8 模型的冬小麦穗识别技术.

Winter Wheat Ear Recognition Based on Improved YOLOv8.

To address the challenges of small target size, dense distribution and occlusion among winter wheat ears in open field environments, this study focused on winter wheat captured by UAV imagery and proposed an improved detection method based on the YOLOv8 model. The SimAM attention mechanism was introduced into the Neck (Neck network) while the GhostNetV2 module was integrated into the C2f module within the Neck. These enhancements improved the representation of spatial and channel features, while maintaining efficient feature fusion and reducing model complexity. As a result, the detection network was better adapted to the complex conditions of open field winter wheat ear detection. In addition, the input image resolution was set to 1280px×1280px to maximize the preservation of critical visual features. The results showed that the improved YOLOv8 model achieved an average precision (AP) of 93.1% and an F1 score of 90.5%, with a model size of only 18.3MB and 9.4 million parameters. Compared to the original YOLOv8, the improved version yield increased of 0.5 percantage point and 0.8 percantage point in AP and F1 score, respectively, while reducing the model size and parameter counted by 3.3MB and 1.7 million parameters. The resulting model is more lightweight and efficient, outperforming the standard YOLOv8 in detecting small, densely distributed and highly occluded winter wheat ears under complex field conditions. [ABSTRACT FROM AUTHOR]

针对大田环境中麦穗目标较小、分布稠密及重叠遮挡等问题, 以无人机拍摄冬小麦为研究对象, 基于YOLOv8 模型提出一种改进的冬小麦穗检测方法, 在 Neck（颈部网络）增加 SimAM 注意力机制, 融合GhostNetV2 模块至 Neck 的 C2f 模块中, 在增强空间和通道特征表达能力、保证特征融合效率的基础上实现了模型轻量化, 使得检测网络更适应复杂的大田环境下麦穗检测, 同时, 设置输入图像分辨率为 1280px×1280px, 最大限度地保留麦穗图像中关键特征信息。结果表明：改进后的 YOLOv8 模型平均精度和 F1 分数分别为93.1%和 90.5%, 权重文件仅占 18.3MB, 参数量 9.4M, 平均精度和 F1 分数较标准 YOLOv8 提高 0.5个和 0.8 个百分点, 权重文件大小和参数量分别降低 3.3MB 和 1.7M, 模型更加轻量化, 整体性能优于原始YOLOv8 模型, 实现了复杂环境下小目标、高重叠度的麦穗数量检测. [ABSTRACT FROM AUTHOR]