PatentOCR: A Benchmark for Fine-Grained Detection and OCR of Component Identifiers in Patent Drawings

Remarkable success has been achieved in natural scenes using optical character recognition (OCR), but challenges in technical domains such as patent drawing remain. Patent drawings contain alphanumeric identifiers that represent patent components, which correspond to the text description in the patent document. However, owing to the various sizes, sparse distributions, and disordered backgrounds of these identifiers, general-purpose OCR engines tend to misidentify components as characters and have difficulty distinguishing relevant identifiers from other texts. Currently, fine-grained annotated datasets tailored for identifiers are lacking. In view of this gap, this study presents PatentOCR, a new dataset containing 2,236 patent drawings, with 20,165 pixel-level fine-grained annotations and their associated arrows or guiding lines, which enables the model to better distinguish target characteristics from complex backgrounds. Moreover, this study leveraged one of the uses of the dataset, i.e., character recognition in patent drawings, established a comprehensive research baseline by evaluating the combination of four models, and introduced two evaluation methods for this task. PatentOCR can provide more resources to promote multimodal research in intellectual property.