Faster R-CNN Based Table Detection Combining Corner Locating
Ningning Sun, Yuanping Zhu, Xiaoming Hu · 2019
Table detection in document images has achieved remarkable improvement. However, there is still a problem of inaccurate table boundary locating. This paper proposes Faster R-CNN based table detection combining corner locating method. Firstly, coarse table detection and corner locating are implemented through Faster R-CNN network. Secondly, those corners belonging to the same tables are grouped by coordinate matching. At the same time, unreliable corners are filtered. Finally, table boundaries are adjusted and refined by corresponding corner group. The proposed method improves the precision of table boundary locating at pixel-level. Experiment results show that our method effectively improves the precision of table detection. It achieves an F-measure of 94.9% on ICDAR2017 POD dataset. Compared with traditional Faster R-CNN method, our method increases by 2.8% in F-measure and 2.1% at pixellevel localization.