Abstract: To achieve real-time detection of tomato at different maturity stages, an improved YOLOv8s-Seg model was proposed to meet the precision management requirements of modern agriculture. By enhancing the neck module of the improved YOLOv8s-Seg model, a 1×1 SimConv layer was added before each upsampling operation, and the remaining conventional convolutions in the neck were replaced with 3×3 SimConv layers, significantly improving feature fusion capability. The results showed that the improved YOLOv8s-Seg model achieved segmentation precision rates of 92.7%, 92.3%, and 89.9% for mature, semi-mature, and immature tomatoes, respectively. Compared with the original YOLOv8s-Seg model, the improved YOLOv8s-Seg model demonstrated increases of 1.6, 0.4, 1.0, and 2.4 percentage points in precision, recall, F1-score, and mAP@0.5, respectively. The improved YOLOv8s-Seg model outperformed YOLOv8s-Seg, YOLOv5s-Seg, YOLOv7-Seg, and Mask R-CNN models in precision, recall, F1-score, and mAP@0.5. The inference time of the improved YOLOv8s-Seg model was 3.5 ms, showing a slight increase compared to YOLOv5s-Seg and YOLOv8s-Seg models, but remained significantly lower than YOLOv7-Seg and Mask R-CNN models.The improved YOLOv8s-Seg model exhibited superior performance in tomato maturity segmentation under complex environments, achieving high precision across scenarios involving leaf occlusion, fruit overlap, lighting variations, and viewpoint changes.

Key words: improved YOLOv8s-Seg model, tomato (Solanum lycopersicum L.), maturity, detection