STUDY, DEVELOPMENT AND CHARACTERIZATION OF DEEP LEARNING MODELS FOR OBJECT DETECTION AND SEGMENTATION IN PRECISION AGRICULTURE

The deployment of autonomous robotic platforms in Precision Agriculture is currently hindered by a fundamental bottleneck: the scarcity of annotated training data required for Deep Learning. This thesis addresses this challenge by investigating a "Zero-Data Regime" which utilizes state-of-the-art Open-World Vision-Language models to bootstrap the phenotypic data generation process. To support this methodology, we developed a bespoke Data Engineering middleware designed to bridge the interoperability gap between raw model inference, human-in-the-loop validation tools and standardized training formats. Using this framework, we conducted a rigorous comparative analysis of three distinct architectural families—YOLO-World, OWL-ViT and Grounding DINO—on a specialized dataset of Lactuca sativa. The experimental results identify Grounding DINO as the Pareto-optimal architecture for the "Cold-Start" problem. It achieved a Zero-Shot Mean Average Precision (mAP@50) of 91.34% without any domain-specific training, significantly outperforming the real-time YOLO-World in recall while maintaining lower latency than the Transformer-based OWL-ViT. Ultimately, this work validates the concept of "Indirect Knowledge Transfer," demonstrating that the semantic reasoning of large foundation models can be effectively distilled into curated datasets, thereby shifting the engineering burden from manual pixel-wise labeling to high-level prompt engineering.