Authors

Abstract

Automated detection of tomato ripeness is crucial for achieving precise harvesting and enhancing agricultural productivity. However, detecting tomatoes in natural scenes poses challenges such as missed detections and false positives due to significant variations in target scale, frequent occlusions, and complex backgrounds. Additionally, existing detection models face limitations when deployed on mobile devices. To address these issues, this paper proposes SFH-YOLOv11, a lightweight detection model based on an improved YOLOv11n. Building upon YOLOv11n, this model achieves lightweight performance while maintaining high accuracy through three key enhancements: introducing an attention mechanism in the backbone network to strengthen feature selection capabilities, designing lightweight convolutional modules to reduce model complexity, and reconstructing the feature pyramid network in the neck to enhance multi-scale feature fusion. Experimental results demonstrate that SFH-YOLOv11 outperforms other algorithms, achieving mAP50 and mAP50-95 scores of 91.8% and 78.2%, respectively—representing improvements of 1.7% and 1.0% over the original model. While enhancing performance, SFH-YOLOv11 reduces the number of parameters, computational complexity, and model size by 37.2%, 15.9%, and 34.5%, respectively, compared to the original model. This research provides effective technical support for lightweight maturity detection tasks in complex agricultural scenarios.

Abstract in Chinese

西红柿成熟度的自动化检测对于实现精准采摘和提升农业生产效率具有重要意义。然而，自然场景下的西红柿图像检测存在目标尺度变化大、遮挡频繁和背景复杂引发的漏检与误检问题，以及现有检测模型在移动端部署的局限性。为此，本文提出一种基于改进YOLOv11n的轻量化检测模型SFH-YOLOv11。该模型在YOLOv11n的基础上，通过在主干网络中引入注意力机制以强化特征选择能力、设计轻量化卷积模块以降低模型复杂度、在颈部网络中重构特征金字塔网络以增强多尺度特征融合能力这3个方面进行改进，使得模型在保持高性能的同时实现轻量化。实验结果表明，SFH-YOLOv11的性能优于其他算法，mAP50和mAP50-95分别达到91.8%和78.2%，相较于原模型分别提升了1.7%和1.0%。在性能提升的同时，SFH-YOLOv11的参数量、计算量和模型大小相较原模型分别下降了37.2%、15.9%和34.5%。本研究为复杂农业场景下的轻量化成熟度检测任务提供了有效的技术支持。