AGMSF-VMUNet: attention-guided multi-scale feature fusion VM-UNet for crop pest detection

Crop pest detection (CPD) in field environments is critical yet challenging due to high pest variability, irregular shapes, and complex backgrounds. Conventional CNNs suffer from limited receptive fields, while Vision Transformers (ViTs) are hindered by quadratic computational complexity, limiting their practical deployment. To address these issues, an attention-guided multi-scale feature fusion VM-UNet (AGMSF-VMUNet) model for CPD is constructed. The model integrates three key components: Multi-Scale VSS (MSVSS) blocks that combine state-space models with multi-scale dilated convolutions to capture both global context and local details; Attention Gate Skip Connections (AGSC) that dynamically recalibrate encoder features to enhance pest-relevant information while suppressing background clutter; and a progressive Multi-scale Feature Fusion (MSFF) module that adaptively integrates features across decoder stages through spatial attention. Extensive experiments on the IP102 dataset demonstrate that AGMSF-VMUNet achieves state-of-the-art segmentation performance with 77.22% mean Intersection over Union (mIoU) and 80.21% Dice Similarity Coefficient (DSC). Ablation studies confirm the synergistic contribution of each component. This work provides an effective solution for intelligent crop pest monitoring in real-world agricultural applications.