| 摘 要: 为实现自然环境下麦穗的快速识别,提出一种改进 YOLOv8n模型,以解决麦穗特征不明显、遮挡等问题。通过引入 RepGhostBottleneck和 RepCoreGhostBottleneck提 升 计 算 效 率,采 用 CARAFE (Content-Aware ReAssembly of FEatures)上采样算子提升精度,并利用 Grad-CAM++热力图增强模型的可解释性。实验结果表明,改进后的 YOLOv8n模型在准确率、召回率和平均精度均值(mAP50)上分别达到了90.9%、84.0%和91.2%。与原始 YOLOv8n模型相比,参数量降低了36.6%,准确率、召回率和平均精度 mAP50分别提升了0.4个百分点、1.1个百分点和1.1个百分点。该模型为麦穗识别提供了一种轻量、准确的解决方案,具有良好应用潜力。 |
| 关键词: 图像识别 深度学习 目标检测 YOLOv8 麦穗 轻量化 RepGhos |
|
中图分类号: TP391.4
文献标识码: A
|
| 基金项目: 2021年度兰州市人才创新创业项目 (2021-RC-47);2022 年度科技部国家外专项目 (G2022042005L);2023 年甘肃省高等学校产业支撑项目(2023CYZC-54);2023年甘肃省重点研发计划(23YFWA0013);2023年甘肃农业大学美育和劳动教育教学改革项目(2023-09) |
|
| Research on a Light weight Wheat Ea rDetection Mode lBased on Reparameterization and Information Aggregation |
|
LI Chengshuai, WEI Linjing
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
729783788@qq.com; wlj@gsau.edu.cn
|
| Abstract: To achieve rapid identification of wheat ears in natural environments, an improved YOLOv8n model is proposed to address issues such as inconspicuous features and occlusion of wheat ear. By introducing RepGhost Bottleneck and RepCore Ghost Bottleneck, computational efficiency is enhanced. The CARAFE upsampling operator is adopted to improve accuracy, and Grad-CAM++ heatmaps are utilized to enhance the model’s interpretability.Experimental results show that the improved YOLOv8n model achieves(mAP50)90.9% in precision, 84.0% in recall,
and 91.2% in mean Average Precision. Compared to the original YOLOv8n model, the number of parameters is reduced by 36.6% , while precision, recall, and mAP50 increase by 0.4 percentage points, 1.1 percentage points, and 1.1 percentage points, respectively. This method provides a lightweight and accurate solution for wheat ear recognition, demonstrating promising application potential. |
| Keywords: image recognition deep learning object detection YOLOv8 wheat ears lightweight RepGhost |
