Abstract: With the advancement of intelligent agriculture and precision pesticide application technologies, the accurate detection of farmland weeds is of great significance for increasing crop yields and reducing pesticide usage. However, existing object detection algorithms still faced considerable challenges in complex farmland backgrounds and the detection of small-scale weed targets. To address these issues, this study proposed FAS-DETR, a farmland weed detection algorithm based on an improved RT-DETR. This method integrated the FasterBlock-SCSA hybrid residual module, the AIFI-AgentAttention agent attention mechanism, and the SSFF multi-scale feature fusion module, optimizing the model architecture in terms of feature extraction, global context modeling, and multi-scale information fusion. Specifically, FasterBlock-SCSA enhanced the model's ability to represent fine-grained textures and key regions, AIFI-AgentAttention leveraged agent tokens to improve global feature aggregation, and the SSFF module further improved the detection performance for weed targets at different scales. Experimental results demonstrated that FAS-DETR outperformed both mainstream YOLO series models and the RT-DETR baseline model in metrics including precision, recall, and mAP₅₀. Specifically, its mAP₅₀ reached 76.4%, representing a 3.1 percentage point improvement over RT-DETR baseline model. These results validated the method's effectiveness for farmland weed detection while maintaining reasonable computational and parameter requirements.

Key words: RT-DETR, weed detection, deep learning, object detection