Ensuring an error-free transcription on a full engineering tags dataset through unsupervised post-OCR methods - Archive ouverte HAL Access content directly
Conference Papers Year : 2023

Ensuring an error-free transcription on a full engineering tags dataset through unsupervised post-OCR methods

Abstract

The digital transformation of engineering documents is an ambitious research topic in the industrial world. The representation of component identifiers (tags), which are textual entities without a language model is one of the major challenges. Most of OCR use dictionarybased correction methods so they fail at recognizing hybrid entities composed by numerical and textual characters. This study aims to adapt OCR results on language-free strings with a specific semantics and requiring an efficient post-OCR correction with unsupervised approaches. We propose a two-step methodology to face the questions of post-OCR correction in engineering documents. The first step focuses on the alignment of OCR transcriptions producing a single prediction refined from all OCR predictions. The second step presents a combined incremental clustering & correction approach achieving a continuous correction of tags' transcriptions relatively to their assigned cluster. For both steps, the dataset was produced from a set of 1,600 real technical documents and made available to the research community. When compared to the best state-of-art OCR, the post-OCR approach produced a gain of 9 % of WER.
Fichier principal
Vignette du fichier
pre_print_2023.pdf (746.09 Ko) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

hal-04093197 , version 1 (09-05-2023)

Identifiers

Cite

Mathieu François, Véronique Eglin. Ensuring an error-free transcription on a full engineering tags dataset through unsupervised post-OCR methods. The 17th International Conference on Document Analysis and Recognition (ICDAR 2023), Aug 2023, San Jose (CA), United States. ⟨10.1007/978-3-031-41734-4_6⟩. ⟨hal-04093197⟩
139 View
127 Download

Altmetric

Share

Gmail Facebook X LinkedIn More