Learning from Motion. A Dynamic Feature Fusion Framework for Robust Agricultural Vision in Blurred Environments

Computer vision for precision agriculture, particularly from unmanned aerial vehicles (UAVs), is frequently hampered by motion blur induced by wind and equipment vibration. Traditional approaches treat blur as noise to be removed or rely on computationally intensive restoration, limiting real-time deployment on edge devices. This study re-conceptualises motion blur not as mere degradation but as a source of learnable features characterising object dynamics. We introduce a Dynamic Fuzzy Robust Convolution (DFRC) module, a plug-in enhancement for detection frameworks, which adaptively fuses multi-scale features with synthetically generated fuzzy cues via a transparency-aware mechanism. A key innovation is a parallel CUDA kernel for efficient non-linear interpolation and rotation of feature tensors, preventing boundary overflow and achieving a 38.4× speedup over CPU implementations. Trained on a purpose-built wheat pest dataset (WheatBlur-3K) with paired clear and synthetically blurred images (including uniform and target-localised blur), our YOLOv11-based model demonstrates robust performance. On blurred test sets, it achieves an mAP@0.5 of 86.4\%, a 26.1\% improvement over the baseline, while maintaining a real-time inference speed of 47 FPS. The framework retains effectiveness in adverse conditions like rain, with performance degradation below 8\%. This work provides a practical, efficient solution for blur-robust agricultural monitoring, shifting the paradigm from blur removal to blur-aware perception. Code and data are available at \url{https://gitcode.com/2401_85342087/yolo11-fuzzy-conv.git}.