AI system spots tiny chili flowers hidden in greenhouse glare
Researchers developed a machine-vision system that reliably detects small chili flowers in industrial greenhouses despite poor lighting and plant overlap—a challenge that has frustrated growers for years. Accurate flower detection unlocks better yield forecasting and harvest timing, potentially cutting waste and improving profitability across the global chili supply chain.
Originaltitel: CF-DETR: a robust transformer-based framework for small-scale chili flower detection in industrial chili production systems
Chili pepper (Capsicum spp.) is a high-value industrial horticultural crop widely utilized in food processing, pharmaceuticals, and natural pigment production. Accurate monitoring of flowering is critical for yield formation, pollination management, and early-stage production forecasting in industrial chili production systems. However, in greenhouse environments, chili flowers typically exhibit small object scale and are affected by issues such as lighting variations and occlusion, which pose significant challenges for reliable visual detection. These factors often result in missed detections and unstable performance in practical phenological monitoring tasks. To address these challenges, this study proposes CF-DETR, a robust transformer-based framework for small-scale chili flower detection. Built upon the RT-DETR architecture, the proposed method introduces an efficiency-optimized FasterNet backbone to enhance fine-grained feature extraction for small targets while maintaining computational efficiency. In addition, a dynamic upsampling mechanism is incorporated to preserve structural details during feature reconstruction, and a Bidirectional Multi-scale Attention Feature Pyramid Network (BiMAFPN) is designed to strengthen cross-scale feature interaction under complex greenhouse backgrounds and occlusion conditions. Experiments conducted on a self-constructed greenhouse dataset demonstrate that CF-DETR achieves a Precision of 94.1%, mAP50 of 83.5%, and mAP50–95 of 64.5%, outperforming the baseline RT-DETR-r18 model. Furthermore, deployment on an NVIDIA Jetson AGX Orin platform achieves real-time inference at 30.65 FPS, validating its practical applicability in edge-enabled agricultural systems. The proposed framework provides a reliable visual sensing solution for small-scale phenology monitoring, enabling intelligent pollination management, early yield prediction, and data-driven decision-making in industrial chili production. This work contributes to the advancement of precision horticulture and the digital transformation of industrial crop production systems.