Attention-Guided Wheat Disease Recognition Network through Multi-Scale Feature Optimization

Accurate and timely detection of wheat diseases remains crucial for sustainable agriculture, particularly in major wheat-producing regions. Wheat diseases pose a significant threat to global food security, need precise and timely detection to promote sustainable agriculture. Existing approaches consistently employ single-scale features with shallow-layered convolutional neural networks (CNNs). To bridge the research gaps, we introduce a novel Multi-Scale Wheat Disease Network (MSWDNet) with feature collaboration for wheat disease recognition supported by a comprehensive dataset collected from wheat fields. This study fills research gaps by introducing a novel technique to improve detection accuracy and promote wheat agriculture. Our network uses multistage architecture with progressive feature fusion, incorporating dilated convolution blocks and efficient channel attention mechanisms to capture both fine-grained details and broader contextual patterns. The custom dataset comprises 3,351 high-quality images across five classes collected under diverse environmental conditions. Through extensive experimentation with various CNN backbones, EfficientNet-B7 emerged as the optimal feature extractor, achieving 92.55% accuracy. Our complete architecture, enhanced with multi-scale feature integration and channel attention mechanisms, achieved 98.50% accuracy. Comprehensive ablation studies validate the effectiveness of each architectural component.