Objektdetektering med AI för projektbaserad tillverkning : RF-DETR vs. YOLOv12

In project-based manufacturing, efficiently organizing and reusing visual material is critical for communication and knowledge retention. This thesis explores whether fine-tuned object detection models—YOLOv12 and RF-DETR—can automate metadata tagging for a Swedish furniture SME that lacks a Digital Asset Management (DAM) system and relies on Dropbox. A Design Science Research (DSR) approach was used. Both models were trained on a 33-class dataset with 9,447 annotated images (including augmentation). YOLOv12 is convolutional and optimized for speed; RF-DETR is transformer-based. RF-DETR showed slightly higher accuracy ([email protected]: 0.82 vs. 0.79), while YOLOv12 was faster (36s vs. 80s per 40-image batch). Both outperformed manual tagging in speed, though occlusion and ambiguity remained challenging in comparison to human-level accuracy. Beyond benchmarking, the study identifies key organizational enablers and constraints for AI adoption, including the need for semantic consistency, responsibility for model maintenance, and alignment with informal practices. The company reflects characteristics of Mintzberg’s Simple Structure—centralized decision-making and informal coordination—yet displays early signs of systematization and potentially more structured operational roles. In this context, the proposed pipeline could enhance DAM capabilities by improving searchability, reducing manual overhead, and enabling cross-project reuse. The thesis contributes both a replicable pipeline and practical insights for implementing sustainable, domain-specific AI in small organizations lacking dedicated AI roles.