Abstract
Tomato growing points and flower buds serve as vital physiological indicators influencing yield quality, yet their detection remains challenging in complex facility environments. This study develops an improved YOLOv8 model for robust flower bud detection by first constructing a comprehensive multi-environment dataset covering 10 typical growing conditions with enhanced annotations. Three key innovations address YOLOv8's limitations: (1) an SE attention module boosts feature representation in cluttered environments, (2) GhostConv replaces standard convolution to reduce computational load by 19% while preserving feature discrimination, and (3) a scale-adaptive WIoU_v2 loss function optimizes gradient allocation for variable-quality data. Ablation experiments confirm these modifications synergistically improve adaptability to scale and environmental variations, achieving 97.8% mAP@0.5 (+ 0.5%) and 85.1% mAP@0.5:0.95 (+ 5.1%) with 11% fewer parameters. Practical deployment on agricultural robots in operational greenhouses demonstrated 93.6% detection accuracy, validating the model's effectiveness for precision agriculture applications. The proposed system achieves an optimal balance of accuracy, speed, and lightweight design while providing immediately applicable solutions for automated tomato monitoring.