基于YC-YOLO v7模型的油菜幼苗株数识别方法

Identification of Rapeseed Seedling Number Based on YC-YOLO v7 Model

摘要:

针对大田环境下油菜幼苗尺度不一、分布密集、识别难度大等问题，开展了基于YC-YOLO v7模型的油菜幼苗株数识别研究。在原模型YOLO v7的高效聚合网络ELAN中引入深度可分离卷积模块，提高模型对细小特征的提取能力；通过在主干网络输出的特征层中添加CBAM注意力机制模块，加强模型对小目标的识别精度；将损失函数CIOU替换为WIOU，提高了锚框质量；为扩大模型对目标的感受野，构建了SPPF空间金字塔结构。试验结果表明，改进后YC-YOLO v7模型平均精度均值为94.0%，精确率为89.8%，召回率为91.2%，推理速度提高16.1f/s，浮点运算量降低2.56×1010；与其他一阶段模型YOLO v5s、SSD和二阶段模型Faster R-CNN进行对比，平均精度均值分别提高12.8、17.8、20.3个百分点。基于YC-YOLO v7模型搭建的油菜幼苗检测识别系统准确率大于90%，可为大田环境下油菜幼苗精准计数提供技术支撑。

Abstract:

In response to problems such as different graphics, densely distributed, and difficult to identify in the field environment, the study of the number of rapeseed seedlings based on the YC-YOLO v7 algorithm was carried out. Introduce the depth-separated convolutional module in the ELAN of the original model YOLO v7 to improve the extraction ability of the model on small features. By adding the CBAM attention mechanism module to the feature layer output by the main network, the model of the models identification of small targets is enhanced. Replace the loss function CIOU to WIOU, which improves the quality of the anchor frame. In order to expand the model of the model for the goal, the SPPF space pyramid structure was constructed. The test results show that the average accuracy of the improved YC-YOLO v7 model was 94.0%, the accuracy was 89.8%, the recall rate was 91.2%, the reasoning speed increased by 16.1.f/s, and the floating-point computing volume was reduced by 2.56×1010. Compared with the other phase model YOLO v5s, SSD, and second-stage model Faster R-CNN, the average accuracy increased by 12.8 percentage points, 17.8 percentage points, and 20.3 percentage points, respectively. The improved YC-YOLO v7 model was deployed to the PC, and an oilseed rape seedling detection and identification system was constructed using the PYQT5 framework, with the average accuracy of the system detection being greater than 90%, which can provide technical support for the accurate counting of oilseed rape seedlings in the field environment, and provide effective support for the farmers to judge the quality of the breeding and the effect of sowing.