PingAn-VCGroup's Solution for ICDAR 2021 Competition on Scientific Literature Parsing Task B: Table Recognition to HTML

Abstract

This paper presents our solution for ICDAR 2021 competition on scientific literature parsing task B: table recognition to HTML. In our method, we divide the table content recognition task into four sub-tasks: table structure recognition, text line detection, text line recognition, and box assignment. Our table structure recognition algorithm is customized based on MASTER [1], a robust image text recognition algorithm. PSENet [2] is used to detect each text line in the table image. For text line recognition, our model is also built on MASTER. Finally, in the box assignment phase, we associated the text boxes detected by PSENet with the structure item reconstructed by table structure prediction, and fill the recognized content of the text line into the corresponding item. Our proposed method achieves a 96.84% TEDS score on 9,115 validation samples in the development phase, and a 96.32% TEDS score on 9,064 samples in the final evaluation phase.

1st Place Solution for ICDAR 2021 Competition on Mathematical Formula Detection

“…• Ranger optimizer has outperformed Adam optimizer consistently. Similar observation is also found in our another report [25] about ICDAR 2021 Competition on Scientific Table Image Recognition to LaTeX [26]. In our evaluation of standard benchmarks, we also find that Ranger can improve the average accuracy by around 0.6%.…”

Rethinking Image-Based Table Recognition Using Weakly Supervised Methods

“…Recently, some researchers (Nassar et al, 2022;Qiao et al, 2021;Ye et al, 2021;Zhang et al, 2022) worked on both table structure recognition and cell content recognition to build a complete table recognition system. J. Ye et al (Ye et al, 2021) (Jimeno Yepes et al, 2021).…”

“…Recently, some researchers (Nassar et al, 2022;Qiao et al, 2021;Ye et al, 2021;Zhang et al, 2022) worked on both table structure recognition and cell content recognition to build a complete table recognition system. J. Ye et al (Ye et al, 2021) (Jimeno Yepes et al, 2021). These two-step approaches achieved competitive accuracies; however, these systems rely on training datasets containing many richly annotated table images.…”

“…In recent years, many table recognition methods have been proposed and proven to be powerful models for both PDF and image-based table recognition. However, most previous studies (Nassar , 2022;Qiao et al, 2021;Ye et al, 2021;Zhang et al, 2022) are based on fully supervised learning approaches. They rely on training datasets that contain lots of richly annotated table images.…”

“…6. García et al, 2020;Huang et al, 2019) and table recognition (Deng et al, 2019;Ye et al, 2021;Zhong et al, 2020), which we will briefly survey here. Most of the previous methods for table recognition are based on two-step approaches dividing the problem into two steps.…”