Tomato Leaf Disease recognition based on Fine-Grained Interpretable Knowledge Distillation model for smart agricultural

In large-scale smart agricultural plantations, in order to utilize computer vision technology for automatic recognition of Tomato Leaf Diseases (TLD) and improve the intelligence level of Smart Agricultural Internet of Things (SAIoT), this paper designs a novel Fine-Grained Interpretable Knowledge Distillation (FGIKD) model. Firstly, based on Deformable Dilated Convolution (DDC) and Simplified Self-Attention (SSA) mechanism, a new Deformable Multi-Scale Perception (DMSP) spatial attention module is designed to integrate the irregular local perception ability of DDC with the global modeling ability of self-attention, thereby enhancing the low-level visual feature extraction capability of the model. Secondly, based on Cross-Layer Feature Fusion (CLFF) and Graph Self-Supervised Learning (GSSL), a new Fine-Grained (FG) feature extraction module is designed to alleviate the problem of "high intra-class variance and low inter-class variance" in TLD images. Thirdly, DMSP and FG distillation functions are designed to transfer the knowledges from teacher network to student network, enabling it to achieve performance close to the teacher network with a small number of parameters. Finally, combining class activation maps with regional confidence weighting technique, a new CNN model post-hoc explanation scheme is designed in the form of "saliency map". In the comparison experiments of standard dataset validation and real-world application testing, the knowledge-distilled student network achieves 98.13 % and 97.56 % TLD recognition accuracies, while the number of model parameters is only 2.921MB, which can meet the requirements of SAIoT terminal model deployment. • Designed FDIKD model can improves the performance of lightweight CNN network. • Designed DMSP can improves model's ability to perceive irregular regions in image. • Designed graph self-supervised learning can refine the fine-grained features of image. • Designed regional confidence weighting can provide visual interpretation for CNN .