Layer-Wise Dual Attention Network with Adaptive Feature Fusion for Apple Disease Classification

This paper proposes a Layer-wise Dual Attention Network (LDA-Net) for apple disease classification, which incorporates both ECA (Efficient Channel Attention) and CBAM (Convolutional Block Attention Module) attention mechanisms in the layer4 stage of ResNet50. The network introduces a novel feature fusion strategy that effectively combines different attention information, along with an enhanced classification head design. Additionally, we adopt Focal Loss to address the challenges of class imbalance and hard example learning in disease classification. Our approach is evaluated on a comprehensive dataset comprising four categories. Experimental results show that LDA-Net demonstrates improved performance with a test accuracy of 99.47%, performing better than traditional models. The model maintains exceptional performance across all disease categories, with precision and recall consistently above 0.99, demonstrating its robustness and reliability for practical apple disease diagnosis applications.