Answers are grounded in this paper and cite their sources.
In greenhouse tomato cultivation, fruit detection and counting are critical tasks for yield prediction and automated harvesting. However, actual greenhouse environments often present challenges such as illumination variations, dense leaf occlusion, and complex background interference. To address these issues, this study proposes a dual-branch fusion detection model guided by high-frequency information (FGfuse-Net). The model introduces high frequency image information through Fourier transform and designs a frequency guided feature fusion module (FGFF) to effectively integrate spatial domain enhanced features, thereby improving the model's adaptability to illumination variations. Meanwhile, a multi-scale wavelet extractor (MSWE) module is introduced to enhance fine-grained texture representation in complex backgrounds via wavelet transform, thus improving detection performance for small and occluded fruits. Furthermore, a Unified-IoU loss function is adopted to optimize bounding box regression, alleviating the gradient imbalance problem in dense scenes and further enhancing detection stability. The model was trained and evaluated on a self-constructed greenhouse tomato dataset. Experimental results demonstrate that it achieves 96.4% mAP@0.5 and stable counting performance. Moreover, it maintains robust performance even under low light conditions, effectively addressing multiple detection challenges in complex greenhouse environments.
No takes yet. Share an insight, caveat, or question.
Tang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: