YOLOv12-BDA: A Dynamic Multi-Scale Architecture for Small Weed Detection in Sesame Fields

Sesame (<i>Sesamum indicum</i> L.) is one of the most important oilseed crops globally, valued for its high content of unsaturated fatty acids, proteins, and essential nutrients. However, weed infestation represents a major constraint on sesame productivity, competing for resources and releasing allelopathic compounds that can significantly reduce both yield and quality without timely control. To address the challenge of low detection accuracy in complex agricultural environments with dense weed distributions, this study proposes YOLOv12-BDA, a dynamic multi-scale architecture for small weed detection in sesame fields. The proposed architecture incorporates three key dynamic innovations: (1) an Adaptive Feature Selection (AFS) dual-backbone network with a Dynamic Learning Unit (DLU) module that enhances cross-branch feature extraction while reducing computational redundancy; (2) a Dynamic Grouped Convolution and Channel Mixing Transformer (DGCS) module that replaces the C3K2 component to enhance real-time detection of small weeds against complex farmland backgrounds; and (3) a Dynamic Adaptive Scale-aware Interactive (DASI) module integrated into the neck network to strengthen multi-scale feature fusion and detection accuracy. Experimental validation on high-resolution sesame field datasets demonstrates that YOLOv12-BDA significantly outperforms baseline models. The proposed method achieves mAP@50 improvements of 6.43%, 11.72%, 7.15%, 5.33%, and 4.67% over YOLOv5n, YOLOv8n, YOLOv10n, YOLOv11n, and YOLOv12n, respectively. The results confirm that the proposed dynamic architecture effectively improves small-target weed detection accuracy at the cost of increased computational requirements (4.51 M parameters, 10.7 GFLOPs). Despite these increases, the model maintains real-time capability (113 FPS), demonstrating its suitability for precision agriculture applications prioritizing detection quality. Future work will focus on expanding dataset diversity to include multiple crop types and optimizing the architecture for broader agricultural applications.