Detection of Melon Leaf Diseases Based on an Improved YOLOv8n Model

Under field conditions, melon leaf diseases suffer from irregular lesion morphology, missed detection of small lesions, insufficient localization accuracy, and considerable difficulty in edge device deployment. Using YOLOv8n as the baseline, this paper performs scenario-adaptive modifications of existing fundamental modules for melon disease detection and proposes the YOLOv8n-melon model. Specifically, an Inner-ShapeIoU loss function is constructed by combining the inner-bounding-box mechanism of Inner-IoU with the shape-aware constraint of Shape-IoU, with the hyperparameters determined via grid search, to improve the regression of irregular lesions. The EMA attention module is appended to the output of each C2f block in the neck, forming C2f-EMA units that suppress complex background interference and strengthen multi-scale lesion feature extraction. A LightHead lightweight decoupled detection head, built with SimAM and depthwise separable convolution (DWConv), is adopted to reduce computational cost. Ablation and comparative experiments are conducted on a melon leaf disease dataset integrated from multiple public sources. The proposed model contains only 2.38 M parameters and 5.9 GFLOPs, achieves an [email protected] of 91.87% with a 2.17-percentage-point improvement over the baseline YOLOv8n (89.70%), and provides a favorable accuracy-complexity trade-off for edge-oriented disease detection compared with mainstream models such as YOLOv3-tiny, YOLOv5n, YOLOv8s, YOLO11n, and RT-DETR‑L. On the Raspberry Pi 4B platform, an inference speed of 10.52 FPS is achieved on static test-set images, verifying the feasibility of edge inference and providing a reference for lightweight crop disease detection.