| 摘 要: 为了快速精准地在复杂果园环境下（如远视距、果叶遮挡和密集果群等场景）对苹果果实进行检测，本文提出了一种基于YOLOv9s的苹果检测模型。首先，将SCSA（Spatial and Channel Synergistic Attention）注意力机制集成到YOLOv9s的模型当中，通过抑制枝叶遮挡和背景等特征信息，使模型更加关注待检测的果实区域；其次，使用提取特征更加高效的DualConv模块对原始模块进行替换，通过多尺度特征融合方法提 高小目标检测能力；最后，使用 Inner-MPDIoU 损失函数对模型进行优化，提高识别精度。实验结果表明：所提出算法在测试集下的准确率、召回率、mAP0.5、mAP0.5～0.95分别达到88.9%、88.5%、93.8%和62.2 %，较基准模型相比均有提升，同时兼顾了良好的实时检测性能。 |
| 关键词: 深度学习 目标检测 YOLOv9s 损失函数 SCSA |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家重大仪器专项(2020YFC2008704) |
|
| Orchard apple detection algorithm based on improved Yolov9s |
|
liuzilong, zhangjindong
|
School of Optical-Electrical and Computer Engineering (OECE) at University of Shanghai for Science and Technology
|
| Abstract: To achieve rapid and precise detection of apple fruits in complex orchard environments (such as long-distance viewing, fruit-leaf occlusion, and dense fruit clusters), this paper proposes an apple detection model based on YOLOv9. First, the SCSA attention mechanism is integrated into the YOLOv9 model to suppress irrelevant features like occluding leaves and background, thereby enhancing the model's focus on target fruit regions. Second, the original convolution module is replaced with the more feature-efficient DualConv module, leveraging multi-scale feature fusion to enhance small-target detection capability. Finally, the model is optimized using the Inner-MPDIoU loss function to further improve recognition precision. Experimental results demonstrate that the proposed algorithm achieves precision, recall, mAP0.5, and mAP0.5～0.95 scores of 88.9%, 88.5%, 93.8%, and 62.2%, respectively, on the test set. The improved algorithm outperforms the original model in most metrics. Comparative experiments confirm that the method exhibits superior robustness while maintaining real-time performance, ensuring both high detection accuracy and practical applicability in orchard environments. |
| Keywords: deep learning object detection YOLOv9s loss function SCSA |
