YOLO-MSS: A lightweight multi-scale detection model for dense multi-species fusarium spores

Fusarium head blight (FHB) poses a serious risk to wheat production worldwide, and accurate spore detection is essential for early warning and precision management. However, microscopic detection remains highly challenging, as spores typically appear extremely small and densely clustered, with adhesion and occlusion in the images, as well as subtle differences between classes and considerable variability within each class, making reliable detection and classification under complex backgrounds particularly difficult. To better reflect these characteristics of Fusarium spores in practical scenarios, a multi-type Fusarium spore dataset containing three representative species, termed the Fusarium-SOD Dataset (Small, Occluded, and Dense), was constructed to capture the small, occluded, and densely distributed characteristics of Fusarium spores, and an efficient detection framework, YOLO-MSS, was developed based on YOLOv8n. The model introduces a Morphology-Density Fusion Convolution (MDFConv) module to enhance feature representation for small and dense targets, integrates a Sequential Fusion Attention (SFA) mechanism into the C2f structure to improve feature discrimination, and adopts a Scale-based Dynamic Loss for the BBox (SDB) loss to stabilise localisation across different scales. Experimental results indicate that YOLO-MSS attains an mAP@0.5 of 97.3%, along with Precision and Recall values of 94.1% and 95.1%, respectively. The model also maintains a lightweight design, with a compact size of 6.15 MB and an inference speed of 94 FPS, thereby meeting the practical requirements for real-time spore monitoring and hardware deployment.