CDPNet: a deformable ProtoPNet for interpretable wheat leaf disease identification

Accurate identification of wheat leaf diseases is crucial for food security, but existing prototype-based computer vision models struggle with the scattered nature of lesions in field conditions and lack interpretability.

To address this, we propose the Contrastive Deformable Prototypical part Network (CDPNet). The idea of CDPNet is to identify key image regions that influence model decisions by computing similarity measures between convolutional feature maps and latent prototype feature representations. Moreover, to effectively separate the disease target area from its complex background noise and enhance the discriminability of disease features, CDPNet introduces the Cross Attention (CA) Mechanism. Additionally, to address the scarcity of wheat leaf disease image data, we employ the Barlow Twins self-supervised contrastive learning method to capture feature differences across samples. This approach enhances the model's sensitivity to inter-class distinctions and intra-class consistency, thereby improving its ability to differentiate between various diseases.

Experimental results demonstrate that the proposed CDPNet achieves an average recognition accuracy of 95.83% on the wheat leaf disease dataset, exceeding the baseline model by 2.35%.

Compared to other models, this approach delivers superior performance and provides clinically interpretable decision support for the identification of real-world wheat diseases in field settings.