Tag recognition from panoramic scans of industrial facilities

Abstract

CAD-based digital twins are commonly used by operators of process industry facilities to combine 3D models with external information and documentation. However, often a suitable model does not exist, and the plant operators must instead resort to laser scans with panoramic photos, which provide little to no metadata or information about their contents. Reading of equipment tags or other useful text from these scans could hugely increase their usefulness, as that information could be used to connect equipment to its documentation and other data. In this thesis, the feasibility of such extraction as a special case of deep learning text detection and recognition is studied.

This work contrasts practical requirements of industry with the theory and research behind text detection and recognition, with experiments conducted to confirm the feasibility of a potential application. It is found that the task is feasible from both business domain and deep learning perspectives. In practice, off-the-shelf text detection models generalize very well to the problem but integrating text recognition to build an end-to-end solution is found to require further work. End-to-end text recognition models appear promising in research, but rather uncommon in practical applications. Recent laser scans including color imagery are found suitable for the task and using them for recognition is found feasible; however, the usefulness of older scans remains unclear due to their poor quality. Deploying a successful practical solution is thus possible with modern scans but acquiring such scans may require collaboration with facility operators.