A natural scene
text recognition method based on geometric prior and a
knowledge graph comprises the following steps: acquiring a field scene text image, detecting and
cutting out a text line image, and performing
feature extraction and columnar deformation correction through a deformation correction model based on geometric prior; perceiving each character of the corrected image through a
visual identification module based on an attention mechanism to obtain aligned
visual texture features of a character level; then, scene
domain knowledge is introduced through a global semantic reasoning module based on
a domain knowledge graph, context information is sensed, and high-level semantic features are coded; and finally, integrating the output of the visual and semantic modules to obtain a
text recognition result. The method can be migrated and applied to different field-oriented natural scene
text recognition of
automatic control instruments, equipment manufacturing,
numerical control machine tools, automobile manufacturing,
rail transit and the like, the problem that the recognition accuracy is not high due to cylindrical text deformation and lack of related dictionaries in natural scenes in a traditional text recognition technology is solved, and more accurate recognition of field texts is achieved.