Detección de objetos de última generación: una revisión del rendimiento y modelos avanzados frente a los estándares en la agricultura inteligente

Precision agriculture incorporates computer vision techniques to automate image analysis and optimize monitoring tasks in production systems. This paper presents a comparison of object detection models applied to a practical precision agriculture case study, using a proprietary dataset of grape bunch images, with a methodology extensible to standard datasets such as COCO. Four representative approaches were evaluated: Faster R-CNN as a classic two-stage model, YOLOv11 as a widely used single-stage detector, and two more recent transformer-based architectures, DINO and D-FINE. The comparison was performed considering accuracy metrics, such as mAP@50, and computational performance metrics, including inference speed. The results show that YOLOv11n offers the best compromise between accuracy and efficiency, outperforming the other evaluated models. Faster R-CNN exhibits lower performance, comparable to that of DINO, while D-FINE achieves higher accuracy at the cost of significantly higher computational resources. These results highlight the importance of selecting the detection model based on the application context and the operational constraints specific to the agricultural environment.