| 摘要: |
| 为构建田间杂交大豆胚轴颜色检测模型，以大田场景下的大豆植株为研究对象，利用自走式大豆表型信息采集平台获取图像数据并构建杂交大豆胚轴颜色数据集，使用不同目标检测模型（SSD、Faster R-CNN、YOLOv3、YOLOv4、YOLOv5、YOLOX和YOLOv7）对杂交大豆胚轴颜色数据集进行检测，将模型分数（F1）、平均精度均值（mAP）及检测速度3个指标用于评估不同模型在杂交大豆胚轴颜色检测中的性能。在YOLOv7网络中添加CARAFE特征上采样算子、SE注意力机制模块和WIoU位置损失函数，建立杂交大豆胚轴颜色检测模型YOLOv7-CSW，并利用改进模型对杂交大豆胚轴颜色数据集进行消融试验。结果表明：1）YOLOv7模型的F1（0.92）与mAP（94.3%）均显著高于其他模型；2）YOLOv7模型的检测速度为58 帧/s，低于YOLOv5和YOLOX，检测速度可以满足田间实时检测任务需求；3）YOLOv7-CSW模型比YOLOv7模型的F1和mAP分别升高0.04和2.6%；4）YOLOv7-CSW模型比YOLOv7模型检测速度升高了5帧/s，可以实现杂交大豆胚轴颜色实时检测。综上，YOLOv7-CSW模型可以更好地获取胚轴颜色特征并准确地检测出目标位置，提高了复杂农田环境下的目标检测性能，能够实现田间杂交大豆胚轴颜色快速准确检测。 |
| 关键词: 大豆 杂种优势 胚轴颜色检测 YOLOv7网络 目标检测 |
| DOI：10.11841/j.issn.1007-4333.2024.02.02 |
| 投稿时间：2023-07-31 |
| 基金项目:国家大豆产业技术体系（CARS-04-PS30）；黑龙江省保护性耕作技术研究中心平台建设（PTJH202102）；黑龙江八一农垦大学研究生创新科研项目（YJSCX2022-Y20） |
|
| Color detection model of hybrid soybean hypocotyl based on an improved YOLOv7 object detection model |
|
YU Chuntao1, LI Jinyang1, SHI Wenqiang1, QI Liqiang1, GUAN Zheyun2, ZHANG Wei1*, ZHANG Chunbao2*
|
| (1.College of Engineering, Heilongjiang Bayi Agricultural University, Daqing 163319, China;2.Soybean Research Institute/Key Laboratory of Hybrid Soybean Breeding of Ministry of Agriculture and Rural Affairs, Jilin Academy of Agricultural Sciences, Changchun 130033, China) |
| Abstract: |
| In order to build a color detection model of hybrid soybean hypocotyl， soybean plants in natural growing environment were used as research objects. This study used a self-propelled soybean phenotyping information collection platform to obtain soybean plant image data in the farmland， and built hybrid soybean hypocotyl color dataset. Color dataset of hybrid soybean hypocotyl were detected with different object detection models （SSD， Faster R-CNN， YOLOv3， YOLOv4， YOLOv5， YOLOX and YOLOv7）. The model score （F1）， mean of average precision （mAP）， and detection speed were compared to evaluate the performance of different models in color detection of hybrid soybean hypocotyl. A color detection model of hybrid soybean hypocotyl， YOLOv7-CSW， was established by adding CARAFE feature up sampling operator， SE attention mechanism module and WIoU box loss function into the YOLOv7 network. The improved model was used to perform ablation tests on the hybrid soybean hypocotyl color dataset. The results showed that： 1） The F1 （0.92） and mAP （94.3%） of YOLOv7 model were significantly higher than those of other models； 2） The detection speed of YOLOv7 model was 58 fps， which was only lower than YOLOv5 and YOLOX， but the detection speed could meet the requirements for real-time detection tasks； 3） Compared with YOLOv7 model， the F1 and mAP of YOLOv7-CSW model were increased by 0.04 and 2.6% respectively； 4） The detection speed of YOLOv7-CSW model was 5 fps higher than YOLOv7 model， and it realized real-time detection of hybrid soybean hypocotyl color. To sum up， YOLOv7-CSW model could better obtain hypocotyl color features and accurately detect the object position， improved the object detection performance in complex farmland environments， and realized rapid and accurate detection of hypocotyl color of field hybrid soybean. |
| Key words: soybean heterosis hypocotyl color detection YOLOv7 network object detection |
