Novel Grad-CAM based Mask-Guided Attention for Plant Disease Classification on Edge

In the context of edge computing, where computational resources and power are constrained, the deployment of machine learning models poses a challenge due to the trade-off between model size and accuracy. Obtaining plant leaf datasets is resource-intensive, resulting in limited availability of data for plant disease classification. To address these challenges, we propose a novel approach called Grad-CAM based Mask Guided Attention (GCB-MGA), which leverages mask-guided attention to transfer knowledge from a large model to a small one by using grad-CAMs from the large model. This methodology enables the small model to effectively learn and mimic the significant features identified by the large model through masks, tackling both model size and data scarcity issues simultaneously. Our experimental results with four different plants demonstrate consistent improvements, with a notable 7% improvement in accuracy for the tomato dataset. Moreover, the proposed GCB-MGA is shown to exhibit superior performance when trained with a quarter of the entire dataset compared to conventional models using the entire dataset, all while incurring negligible hardware overhead. We demonstrate that the proposed GCB-MGA achieves an inference time of 12 ms with an energy consumption of 61.48 mJ per inference on Jetson Nano, ensuring negligible power overhead.