While image-based georeferencing systems are widely available, none are designed for edge computing on board small satellites in space. Recent advances in autonomous data processing allow points of interest (POIs) to be identified in the captured images, enabling prioritization of data and reducing required downlink bandwidths. Directly attributing geopositional information to these POIs distills the generated insights to labeled locations, eliminating the need for images to be transferred off the satellite platform. Current georeferencing pipelines rely on databases of reference images and substantial computational resources to extract and match keypoints; an operation not feasible with the raw data and hardware limitations on board low-powered satellites. This work proposes a novel architecture which addresses these challenges. By pre-extracting reference keypoints using terrestrial resources, minimal processing on board is reserved for the captured images. Using a modular system architecture, the design is generalized for various small satellite platforms. A proof-of-concept analytical model is implemented based on D2-Net. This is evaluated on a variety of Sentinel-2A datasets, and the resulting state-of-the-art spatial accuracy confirms the feasibility and significant potential of this architecture. Finally, the limitations of this implementation are explored, from which specific recommendations for future improvements are proposed.