Abstract: To improve the detection accuracy and speed of potato leaf diseases and address false detections and missed detections caused by complex environments and diverse lesion types, a potato leaf disease detection model based on multi-path parallel downsampling, named YOLOv11-RFD, was proposed. The dataset was constructed using the publicly available PlantVillage dataset and the Potato Leaf Disease Dataset. Potato leaf images were extracted and merged into a self-built dataset containing three categories of images captured under both single backgrounds and real-field environments. To improve network performance, the downsampling layers of YOLOv11n were replaced with the RFD module, which integrated multiple downsampling methods to generate more robust multi-dimensional feature representations and enabled the model to effectively capture critical information in the shallow stages of the network. Detection results on the self-built dataset showed that the YOLOv11-RFD model achieved a precision of 94.8%, an mAP_0.5 of 95.4%, and an mAP_0.5:0.95 of 88.0%. Compared with the baseline model YOLOv11n, the improved network enhanced the discriminative ability for disease region features, resulting in a 7.6 percentage point increase in precision. However, as the feature screening mechanism was strengthened, some low-saliency, small-scale, or boundary-blurred lesions were suppressed, causing a small number of real targets to remain undetected and leading to a certain decrease in recall. Meanwhile, mAP_0.5 and mAP_0.5:0.95 increased by 1.6 and 0.5 percentage points, respectively. The results demonstrated that the YOLOv11-RFD model could effectively improve overall detection performance and provided a reference for the practical application of potato leaf disease detection.

Key words: potato leaf disease, YOLO, object detection