Cross‐Scene Robustness Study of Transformer‐Based Tomato Fruit Ripeness Detection Under Complex Illumination Conditions

Reliable tomato ripeness detection under complex illumination is important for intelligent harvesting, automated grading, and greenhouse production management. However, illumination changes in facility agriculture, such as low light, strong reflection, supplemental LED lighting, and color‐temperature shifts, can cause obvious cross‐scene performance degradation. These disturbances also increase confusion between the turning class and adjacent ripeness categories. To address this problem, this study proposes TEC‐DETR, an improved tomato ripeness detection model based on RT‐DETRv2‐R50. Three modules are progressively introduced. Training and Data Organization Enhancement (TDOE) improves the coverage of illumination‐related variations during training. Exponential Moving Average Weight Stabilization (EMA‐WS) maintains a smoothed representation of model parameters. Class Discrimination Optimization (CDO) strengthens the learning of easily confused ripeness classes, especially the turning class. A unified evaluation protocol based on the official COCO API and COCOeval was used to compare model performance across the Normal domain and four complex illumination domains. Compared with RT‐DETRv2‐R50, TEC‐DETR increased AP@0.5:0.95 by 0.095, 0.052, 0.0419, and 0.180 on the Low, Strong, LED, and Color_Temp domains, respectively. The mean AP@0.5:0.95 across the four complex domains increased from 0.3615 to 0.4537, and the average cross‐scene performance gap was reduced from −0.3845 to −0.3103. These results indicate that TEC‐DETR improves the cross‐scene robustness of the RT‐DETRv2‐R50 baseline under complex illumination conditions.