IFMA-Net: Interactive Feature Modulated Attention Network for Accurate Identification of Cotton Pests and Diseases from Field Images

Recognizing cotton pests and diseases from field-captured images remains challenging due to subtle symptoms, complex backgrounds, high intra-class variation, and considerable variation in symptom scale. This study proposes IFMA-Net, a ResNet-50-based network that integrates local feature enhancement, multi-scale contextual modeling, and interactive feature fusion for agricultural vision sensing. The Dynamic Local Enhancement (DLE) module enhances discriminative shallow features, while the Global Multi-Scale Attention and Local Aggregation (G-MSLA) module combines multi-scale local aggregation with global dependency modeling. The Interactive Cross-Attention Fusion (ICAF) module employs two opposing cross-attention branches to exchange information between local and global representations. On a self-constructed eight-class field dataset, IFMA-Net achieved 95.53 ± 0.17% accuracy and 94.81 ± 0.14% macro-F1, exceeding ResNet-50 by 5.52 and 7.32 percentage points. Ablation experiments confirmed the individual contributions of the three modules. On an independent public test set, IFMA-Net achieved 90.21% macro-F1. Perturbation experiments showed higher retention than ResNet-50 under random occlusion but substantial degradation under severe Gaussian noise. These results demonstrate IFMA-Net’s potential for accurate pest and disease identification in practical agricultural sensing scenarios.