Abstract: In order to achieve fast and accurate identification of malignant weeds in agricultural fields using deep learning models, this study took two common invasive plants of Erigeron L. in the fields, collected sample images and labeled weed targets, and built a training platform and an embedded test platform based on YOLOv5 with adjustable depth and width of network structure. Fourteen groups of model weights with different network layers and convolution kernels were trained, and the accuracy and detection frame rate were evaluated. The results showed that the average precision of YOLOv5 with different network structure depth and width settings for identifying invasive plants ranged from 91.8% to 95.1%. Eight groups of weights achieved higher average precision than YOLOv3, indicating that the increase of network layers and convolution kernels can improve the model accuracy. The results also showed that frame rates of YOLOv5 were between 28 to 109 fps on the training platform and between 12 to 58 fps on the test platform. Twelve sets of weights exhibited significantly improved frame rates compared to YOLOv3. The frame rate is limited by the computing capability of the platform and decreases with the increase of network layers and convolution kernels. To realize real-time detection in the embedded system with low compute capability, a balanced network structure setting is needed. The results of this study can provide a reference for building an intelligent weed sensing system in agricultural fields.