Information Extraction from Historical Semi-Structured Handwritten Documents

In this paper, we describe our approach to extract salient events such as birth and death records from historical French parish documents that contain free-form handwritten text. The challenges posed by these documents to the current state of the art in handwriting recognition and information extraction go well beyond the generic challenges in recognizing handwritten text such as style variations, irregular baselines, poor legibility, etc. Our approach for extracting salient events from such documents has the following processing steps: (1) pre-processing for noise removal and high-quality binarization, (2) OCR for text recognition, and (3) statistical information extraction for event record extraction. In this paper, we focus on preprocessing techniques for robust binarization in presence of different types of degradations that are common in historical documents. We provide a detailed description of our system, experimental setup, and results for each stage of the processing. In addition, we compare different approaches for preprocessing by assessing their impact on OCR performance.