Leveraging semi-supervised learning and domain generalization techniques for unseen data

This thesis tackles the challenge of domain generalization in object detection, a process demanding abundant annotated data. The focus is on employing semi-supervised learning techniques to streamline annotation, addressing domain shifts in agriculture. By integrating these techniques with models like YOLO and pre-trained model, the study enhances annotation efficiency, fine-tunes model performance, and reduces reliance on extensive labeled datasets. This comprehensive strategy aims to effectively confront domain generalization challenges and reshape applications in sectors like agriculture. At the core of this research lies the intricate task of object detection, delving into the complexities of domain generalization. The integration of semi-supervised learning combines supervised and unsupervised methods, interwoven with various pre-trained models, with YOLO and pre-trained model in the forefront. By skillfully blending these techniques, the research propels annotation processes, optimizes model efficacy, and reduces the need for vast labeled datasets. Beyond traditional approaches, this study ventures into zero-shot object detection with CLIP. Rigorous experimentation and meticulous comparisons, guided by COCO API metrics, underscore this exploration. The thesis emphasizes the urgency of addressing domain generalization obstacles and highlights the transformative potential of semi-supervised learning. This fusion of innovation and practicality not only revolutionizes object detection but also holds promise for driving advancements in diverse sectors, particularly within the intricate fabric of modern agriculture.