Automatic Recovery of Corrupted Font Encoding in PDF Documents Using CNN-based Symbol Recognition with Language Model

In this paper, we present a novel solution for the problem of text extraction from PDF documents with incorrect encoding of embedded fonts. Instead of using a full-scale optical character recognition (OCR) for such documents, our system detects problematic fonts and focuses on their restoration, rather than the actual text. It parsimoniously uses a convolutional neural network (CNN) for performing optical character recognition (OCR) of individual glyphs, classifying them to homoglyph groups. Font glyph-character mappings are recovered by combining glyph location information with a language model. We compare our approach against a fullscale OCR baseline on a new dataset of multilingual PDF files that we made available to the research community. Our evaluation on a variety of test cases shows that our approach has a consistently lower error rate and a significantly faster processing time.