SCUT-CAB: A New Benchmark Dataset of Ancient Chinese Books with Complex Layouts for Document Layout Analysis

Abstract

AbstractAncient books are the cultural heritage of human civilization, among which there are quite a few precious collections in China. However, compared to modern documents, the absence of large-scale historical document layout datasets makes the digitalization of ancient books still in its infancy and awaiting excavation and decryption. To this end, this paper proposes a large-scale dataset named SCUT-CAB for layout analysis of ancient Chinese books with complex layouts. The dataset is established by manually annotating 4000 images of ancient books, including 31,925 layout element annotations, which contains different binding forms, fonts, and preservation conditions. To facilitate the multiple tasks involved in document layout analysis, the dataset is segregated into two subsets: SCUT-CAB-Physical for physical layout analysis and SCUT-CAB-Logical for logical layout analysis. SCUT-CAB-Physical contains four categories, whereas SCUT-CAB-Logical contains 27 categories. Furthermore, the SCUT-CAB dataset comprises the labeling of the reading order. We compare various layout analysis methods for SCUT-CAB, i.e., methods based on object detection, instance segmentation, Transformer, and multi-modality. Extensive experiments reveal the challenges of layout analysis for ancient Chinese books. To the best of our knowledge, SCUT-CAB may be the first large-scale public available dataset for ancient Chinese document layout analysis. The dataset will be made publicly at https://github.com/HCIILAB/SCUT-CAB_Dataset_Release.KeywordsDatasetAncient booksDocument layout analysis