Apple Leaf Disease Analysis Using Deep Hybrid Attention Network

Apple leaf disease (ALD) poses significant challenges to orchard management, affecting both yield and fruit quality. Early and precise detection is important for effective disease management. Moreover, Squeeze-and-Excitation networks (SENets) significantly enhance convolutional neural networks (CNNs) by leveraging channel-wise dependencies within feature maps. This paper introduces a novel approach for automated ALD analysis using a deep hybrid attention network (DHAN) constructed with CNNs and SENets. The CNN component usually captures spatial features from leaf images, while the SENet attention component leverages channel dependencies within the feature representations. We apply this feature re-calibration, generated from the SENet-based attention mechanism, among all feature maps, including residual or identity connection. The extensive experiment shows that our proposed DHAN modifications for 44 layers with widening factors 1 and 2 achieve 100% and 99.84% validation accuracy for 2 and 13-classes datasets, respectively. They consume 0.58M and 2.2M parameters, and 6.3M and 18.4M FLOPS. These performances demonstrate the superiority of the proposed DHAN approach to state-of-the-art methods in terms of apple leaf disease detection.