| 摘 要: 为了提升不同环境中竹笋细粒度的自动化识别精度,提高生产管理效率,文章提出了一种基于YOLOv8的目标检测改进模型。该模型融合了BiFPN (Bidirectional Feature Pyramid Network),这一架构在图像目标检测和分割任务中表现出色,同时在C2f模块中添加DAT(Vision Transformer with Deformable Attention),引入了可变形注意力机制,进一步提升了模型的性能。实验结果表明,改进后的算法模型对春笋和冬笋识别的平均精度均值(Mean Average Precision,mAP)分别为81.4%和94.7%,相较于原有模型,分别提升了0.9百分点和3.9百分点。改进后的算法模型在竹笋细粒度识别方面展现出较高的精度,为未来竹笋产业的高度智能化提供了技术支撑。 |
| 关键词: 细粒度识别;目标检测;可变形注意力机制;多尺度特征;YOLOv8 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 自然科学基金-甘肃省科技计划资助(24JRRA656) |
|
| Fine-grained Recognition Technology of Bamboo Shoots Based on Deep Learning |
|
DUAN Chengxin, ZHAO Xia
|
(College of Inf ormation Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
1121606584@qq.com; 58892778@qq.com
|
| Abstract: In order to improve the accuracy of automated fine-grained recognition of bamboo shoots under various environments, and enhance production management efficiency, this paper proposes an improved target detection model based on YOLOv8. This model integrates BiFPN ( Bidirectional Feature Pyramid Network), an architecture that performs exceptionally well in image object detection and segmentation tasks. Additionally, Vision Transformer with Deformable Attention (DAT) is added to the C2f module to further enhance the model's performance. Experimental results show that the improved algorithm achieves a mean Average Precision (mAP) of 81.4% for spring bamboo shoots and 94.7% for winter bamboo shoots, representing improvements of 0.9 percentage points and 3.9 percentage points respectively, compared to the original model. The improved algorithm model demonstrates high accuracy in finegrained recognition of bamboo shoots, providing a technological foundation for the high automation of the bamboo shoot industry in the future. |
| Keywords: fine-grained recognition; target detection; DAT; multi-scale features; YOLOv8 |
