孙丙宇,单超,房永峰.基于改进YOLOv7的番茄果实目标检测[J].安徽建筑大学学报,2024,32(2):67-72 |
基于改进YOLOv7的番茄果实目标检测 |
Tomato Fruit Target Detection Based on Improved YOLOv7 |
|
DOI: |
中文关键词: YOLOv7 注意力机制 Soft-NMS BiFPN |
英文关键词: YOLOv7 attention mechanism Soft-NMS BiFPN |
基金项目:中国科学院合肥物质科学研究院院长基金重点支持项目(YZJJZX202013) |
|
摘要点击次数: 536 |
全文下载次数: 0 |
中文摘要: |
针对农业采摘机器人在采摘过程中面临果实重叠、果实遮挡和果实体积小难以识别等一系列问题,提出一种改进YOLOv7网络对番茄果实进行目标检测。首先在YOLOv7网络结构中增加SimAM注意力模块和CA注意力模块,提高网络特征提取能力;其次结合特征融合网络的张量拼接操作与加权特征金字塔,提高特征融合能力;再用Soft-NMS算法代替NMS算法,增加网络对重叠区域的检测能力;最后将CIOU Loss替换成EIOU Loss,优化网络性能。实验结果表明,改进后的 YOLOv7网络 mAP值可达 96.7%,准确率为 96.2%,召回率为 99.0%,满足网络对番茄检测精度的要求。 |
英文摘要: |
To solve fruit overlap, occlusion and recognition difficulty caused by small-size fruit for agricultural picking robots, an im‐proved YOLOv7 network was proposed for tomato fruit target detection. Firstly, SimAM and CA attention modules were added to YO‐LOv7 network structure to improve the feature extraction capability. Secondly, the tensor splicing operation of the feature fusion net‐work and the weighted Bidirection Feature Pyramid Network were combined to improve the feature fusion capability. The NMS algo‐rithm was replaced by Soft-NMS algorithm to increase the detection ability in the overlapping area. Finally, CIOU Loss was replaced byEIOU Loss to optimize network performance. The results showed that the improved YOLOv7 network mAP value reached 96.7%; the ac‐curacy reached 96.2%; the recall rate reached 99.0%, which met the network requirements for tomato detection accuracy. |
查看全文
查看/发表评论 下载PDF阅读器 |
关闭 |