Lightweight weed detection method for sugar beet fields using YOLO11-RFL model

Precise and efficient weed detection and removal can be expected to enhance crop growth for the high yields in modern agriculture. It is often required for the high identification speed and accuracy to distinguish sugar beet seedlings from weeds in complex field environments. In this study, a detection model was proposed using an improved YOLO11n architecture—YOLO11-RFL (YOLO11 with Reparametrized Ghost-ELAN, Feature-Fused-Pyramid-Neck, and Lightweight GN head). Firstly, a reparametrized Ghost-ELAN (RG-ELAN) module was used to replace YOLO11n's C3k2 component, according to the inspiration from GhostNet. Reparameterization techniques were employed at module branches to enhance feature extraction with less computational load. Secondly, the Rethinking Features-Fused-Pyramid-Neck (RFPN) framework was proposed to refine the Neck network. Real-time performance and efficiency were enhanced to avoid feature misalignment in multi-scale object detection. Thirdly, GroupNorm was employed to refine the convolutional layer in the detection head. The Lightweight Shared Convolutional Separator GN Detection Head (LSCSGND) was designed to enhance small object localization and classification, thus enabling more accurate weed detection. Finally, a knowledge distillation (teacher-student model) framework was introduced using a feature-based Channel-Wise Knowledge Distillation (CWD). Synergistic optimization was also achieved in the computational efficiency and accuracy for YOLO11-RFL. Experimental analysis was also conducted on the public LincolnBeet dataset. As such, YOLO11-RFL was achieved in the high network performance, compared with the mainstream YOLO models. The precision and recall reached 80.7% and 74.9%, respectively, compared with the rest of YOLO11n. The mean average precision (mAP) at 0.5(mAP_0.5) and 0.5-0.95(mAP_0.5-0.95) was 80.7% and 56.9%, respectively. The performance of YOLO11-RFL was evaluated on the PDT and CWC datasets. The mAP_0.5 was improved by 0.7 and 0.1 percentage points over YOLO11n, respectively, indicating the robustness under diverse environments. Only 2.03 million parameters were required for 5.4 G of computational resources, which was reduced by 21.3% and 14.3%, respectively, compared with the YOLO11n, indicating excellent engineering deployment. Field deployment and performance tests were conducted to validate the inference efficiency and deployment of the YOLO11-RFL model at an experimental base. A real-time detection frame rate of 107.50 frames per second was achieved in a TensorRT inference deployment, indicating the outstanding real-time inference performance and high computational efficiency in practical settings. Additionally, 296 real-world test samples were randomly collected from multiple representative areas in the experimental base. The high performance of the detection was achieved in the baseline model YOLO11n, with an mAP_0.5 of 80.4% on the test set. Whereas the YOLO11-RFL model improved the accuracy to 82.1% in field environments. Furthermore, visualization analysis of real-world images confirmed that the YOLO11-RFL model effectively mitigated accuracy degradation and object omission under environmental complexity. This finding can provide technical support for the precise identification and efficient control of weeds in sugar beet fields. The significant implications can also help advance intelligent agricultural production.