A survey of historical document image datasets

Abstract This paper presents a systematic literature review of image datasets for document analysis, focusing on historical documents, such as handwritten manuscripts and early prints. Finding appropriate analysis is crucial prerequisite to facilitate research using different machine learning algorithms. However, because the very large variety actual data (e.g., scripts, tasks, dates, support systems, amount deterioration), formats label representation, evaluation processes benchmarks, finding difficult task. work fills this gap, presenting meta-study existing datasets. After selection process (according PRISMA guidelines), we select 65 studies that are chosen based factors, year publication, number methods implemented in article, reliability algorithms, dataset size, journal outlet. We summarize each study by assigning it one three pre-defined tasks: classification, layout structure, or content analysis. present statistics, type, language, input visual aspects, ground truth information every dataset. In addition, provide benchmark tasks results from these papers recent competitions. further discuss gaps challenges domain. advocate providing conversion tools common COCO format computer vision tasks) always set metrics, instead just one, make comparable across studies.