Table Information Extraction Using Data Augmentation on Deep Learning and Image Processing

Generally, the extraction of information in the table is done quickly if the table is within a document with a tabular structure. However, in the case of tables presented in the image document, steps are needed first to detect the table. Seeing a table in image documents becomes more complicated if the table to be seen does not have clear boundaries. This research focuses on extracting information from borderless tables in image documents. The study applies the Mask RCNN-FPN deep learning model to detect borderless tables using augmentation data. The use of data augmentation is expected to increase the accuracy of deep learning models even though there is only a small amount of training data available. The data augmentation technique proposed in this study is a fine-tuning method with CutMask augmentation data. For model formation and testing, this study uses the UNLV data set. This data set consists of scanned images of documents from various sources, including financial reports, journals, and different tabular research papers. The total amount of data used is 427 samples. After data augmentation, the amount of data used is 854 samples. The table detection model is based on the Mask RCNN created using Python. The testing parameters used for table detection quality are precise detection, partial detection, false detection, Precision, Recall, and F-Measure. The table's structure recognition rate is measured from the detection intersection value, rows, columns, and cells compared to ground truth. The test results show that using data augmentation with the CutMask technique can improve the performance of deep learning models to detect borderless tables. The use of image processing is shown to enhance table segmentation. However, the table structure recognition result does not offer a good result compared to the effects of other research.