Lightweight Plant Disease Detection With Adaptive Multi-Scale Model and Relationship-Based Knowledge Distillation

Abstract

Plant disease detection is able to control disease spread and help prevent significant food production losses. However, existing detection methods are still limited to different target scales and high model parameters. To this end, we develop a novel framework, that is, FPDD-Net, for lightweight plant disease detection. It is based on YOLOv8 with an adaptive multi-scale model (AMSM) and relationship-based knowledge distillation (RKD). More specifically, the original cross stage partial (CSP) bottleneck is replaced by an AMSM to effectively fuse the multi-scale features. Next, an Alpha-IoU loss optimization is adopted for aligning predicted boxes more precisely with ground truth, leading to fewer localization errors. Finally, RKD is introduced to assist the training and further improve the performance of target detection. To evaluate our network, the FPDD-Net is trained and tested on two typical datasets, that is, the plant village dataset and the plant-doc dataset. Experimental results indicated that our FPDD-Net is lightweight and has advantages over peer methods.