MOIR-Net: A Lightweight Crop Disease and Pest Recognition Network Based onMulti-Scale Reparameterization and Energy Attention

Abstract

Accurate crop pest recognition is critical for intelligent agriculture, yet it remains constrained by the inherent trade-off between complex field variability and the stringent resource limitations of edge devices. In this study, MOIR-Net, a lightweight framework, is proposed to address this bottleneck by integrating Multi-scale Omnikernel Inverted Residuals (MOIR) with Parameter-free SimAM Attention. The MOIR module employs a structural re-parameterization strategy that transforms multi-branch training topologies into single-path inference architectures, thereby enabling lossless feature aggregation with zero inference-time memory overhead. Extensive evaluations conducted on the IP102, PlantVillage, and PlantDoc datasets demonstrate thatMOIR-Net achieves a strong accuracy–efficiency balance while maintaining minimal computational cost (0.967 M parameters, 2.30 ms latency). On the challenging IP102 benchmark, a Top-1 accuracy of 64.18% and mAP of 58.32% are achieved, significantly surpassing state-of-the-art baselines such as MobileViT-v2. Furthermore, the model exhibits superior cross-domain generalization (41.9% Macro-F1 gain) and highly competitive prediction reliability (ECE = 0.0220) through label smoothing. MOIR-Net successfully reconciles extreme computational efficiency with high trustworthiness, thereby offering a robust solution for real-time crop monitoring on resource-constrained agricultural edge platforms.