| Title |
Improving Accuracy of Table Detection in Document Image using Loss Compensation Faster R-CNN |
| Authors |
경민영(Minyoung Kyoung) ; 이현빈(Hyunbean Yi) |
| DOI |
https://doi.org/10.5573/ieie.2021.58.6.61 |
| Keywords |
Table Detection; Deep Learning; Faster R-CNN; Loss Compensation Training |
| Abstract |
Table detection is an important task in document analysis as tables often present essential information in a structured way. But it is difficult to automatically detect tables because their layouts and formats vary. Recently, large amounts of table information and deep learning models have significantly improved table detection performance. However, mis-detection errors which recognize graphs or figures as tables still exist. These errors occur because feature vectors for the table area, extracted by the model, include some feature vectors of the other objects. To reduce these shared feature vectors, we make use of Loss Compensation Training. In this work, we propose a Loss Compensation Module, which defines table and background areas as domains and classifies the two domains, and add our approach to Faster R-CNN so that the intrinsic characteristics of the table can be extracted. Our experimental results demonstrate that the approach significantly reduces the table mis-detections without additional labels for the other objects and improves the performance of table detection compared with Faster R-CNN. |