FEPL-DETR: an efficient lightweight detector for potato leaf diseases under complex field conditions

Potato foliar diseases pose a serious threat to crop growth and yield, making timely and accurate disease detection essential for early intervention and effective field management. However, reliable detection under complex field conditions remains challenging. To improve the reliable localisation of potato leaf lesions under challenging field conditions, this study proposes a lightweight Faster Effective Potato Leaf Disease DETR (FEPL-DETR) based on an improved efficient detection Transformer (RT-DETR). Firstly, we construct the Faster Frequency-domain Dynamic Convolution CSPDarknet (FFDC-CSP) backbone network, enhancing its ability to adaptively represent lesion edges, textures, and irregular boundaries. Secondly, we build a BF-RepELAN feature network by incorporating RepNCSPELAN4 into a bidirectional multi-scale fusion topology. Finally, the DSI-GAM semantic distillation framework is adapted to FEPL-DETR to strengthen semantic supervision during training. Experimental results show that FEPL-DETR achieves 90.6% mAP 50 and 65.4% mAP 50:95 on the DLOP dataset, with 6.5 M parameters, 15.6 GFLOPs. Compared with RT-DETR-R18, FEPL-DETR improves mAP 50 and mAP 50:95 by 5.7 and 0.6% points, respectively, while reducing the number of parameters and computational cost by 67.0% and 72.6%, respectively. Furthermore, FEPL-DETR achieves 13.1 FPS when deployed on an NVIDIA Jetson Orin Nano, demonstrating its applicability to resource-constrained edge devices. External-dataset evaluations further demonstrate its generalisation capability. The proposed model offers an efficient approach to detecting potato leaf diseases under complex field conditions.