Lightweight Weed Detection Model in Complex Agricultural Environments

To address the challenges of small-target weed detection in intelligent weeding robots, this paper proposes a lightweight enhanced YOLOv8-based weed detection model. The improvements specifically target critical issues in imagebased object detection, including the high density and uneven distribution of these weeds, which significantly compromise the practical effectiveness of existing detection methods. First, the P4 and P5 layers are removed, and a new P2 layer is introduced to enhance the model's recognition capability for these weeds, thereby significantly improving small-object detection performance. Additionally, the original C2f module is replaced with a proposed C2f-FC module, which better preserves fine-grained details of small targets while reducing parameter redundancy by 12.6%. Next, an Efficient MultiScale Attention (EMA) mechanism was subsequently integrated between the prediction head and Spatial Pyramid Pooling Fast (SPPF) module to model pixel-wise spatial dependencies. This architectural enhancement enables cross-scale feature interaction by dynamically recalibrating channel-wise attention weights across parallel branches, thereby improving the network's capacity to capture fine-grained contextual relationships in agricultural imagery. Finally, a dynamic Compressed Activation (CA) loss function, specifically engineered for small-target detection, is proposed to accelerate model convergence and enhance detection performance. Experimental results show that the optimized lightweight model achieves 2.0%, 3.2%, and 2.7% improvements in precision, recall, and Average Precision (AP) for small-target weed detection respectively, while reducing network parameters by 40%.