To allow the publication of patents by the European Patent Office, Luminess is responsible for standardizing patent applications received in highly heterogeneous formats into standardized XML and PDF documents. These documents are then made accessible to the scientific community and can be easily searched and retrieved thanks to precise and standardized annotation of their content. However, recognition errors can affect these annotations and prevent relevant documents from being found or retrieved.

This thesis focuses on the automation of mathematical expression recognition in patent applications. These expressions are particularly challenging to recognize automatically and are currently transcribed and manually verified several times to meet the quality requirements of the European Patent Office. The objective of this thesis is thus to reduce the workload associated with their transcription and verification while maintaining a quality level of at least 99.99%.

The first contribution consists of a multimodal evaluation framework for analyzing the performance of optical character recognition systems according to different representations of mathematical expressions. This evaluation highlights the limitations of existing approaches on patent documents and leads to the creation of PatentME-600K, a dataset containing more than 600,000 mathematical expressions extracted from patents. Using this dataset, MML-Net, a Vision Transformer encoder-decoder architecture, is proposed for the recognition of mathematical expressions in MathML. The model achieves a Levenshtein similarity score of 93.7% and outperforms state-of-the-art methods on this task.

To enable the integration of MML-Net into an industrial production pipeline, several complementary components are developed. A scheduler based on a Random Forest and visual features selects the expressions for which the model is most likely to produce a correct transcription, allowing human intervention to be reserved for more complex cases. A post-OCR verifier, based on a Siamese architecture and specifically trained on MML-Net errors, then automatically detects potentially incorrect predictions and routes them for human validation. Finally, a final contribution focuses on assisting operators during transcription and reducing errors introduced during manual operations. Different strategies for integrating these components into Luminess production pipeline are investigated in order to combine automatic recognition and human validation, increase the level of automation, and preserve the quality requirements of the industrial process.