A Benchmark of Parsing Vietnamese Publications

Abstract

In recent decades, digital transformation has received growing attention worldwide, that has leveraged the explosion of digitized document data. In this paper, we address the problem of parsing publications, in particular, Vietnamese publications. The Vietnamese publications are well-known with high variant, diverse layouts, and some characters are equivocal in the visual form due to accent symbols and derivative characters that pose many challenges. To this end, we collect the UIT-DODV-Ext dataset: a challenging Vietnamese document image including scientific papers and textbooks with 5,000 fully annotated images. We introduce a general framework to parse Vietnamese publications containing two components: page object detection and caption recognition. We further conduct an extensive benchmark with various state-of-the-art object detection and text recognition methods. Finally, we present a hybrid parser which achieves the top place in the benchmark. Extensive experiments on the UIT-DODV-Ext dataset provide a comprehensive evaluation and insightful analysis.