Tomato instance segmentation is a fundamental task in precision agriculture, supporting important applications such as yield estimation, disease monitoring, and automated harvesting. Unlike previous studies that mainly focused on traditional CNN-based approaches, this study emphasizes the novelty of integrating lightweight YOLOv12-Seg architectures with newly designed components, specifically R-ELAN and Area Attention, to achieve both high accuracy and real-time inference capability. A dataset of 4,000 tomato images at various growth stages was constructed and carefully annotated to ensure rigorous evaluation. Extensive experiments show that YOLOv12n-Seg significantly outperforms traditional CNNs and achieves competitive accuracy compared to advanced segmentation models, while maintaining low computational complexity suitable for deployment on edge devices and robotic platforms. Statistical validation further confirms the reliability of the results. These contributions highlight the outstanding advantages of YOLOv12-Seg as a practical solution for modern agricultural automation.