A Gradient heatmap based Table Structure Recognition

Most methods to recognize the structure of a table are to use the object detection approach to directly locate each cell in the table or to segment the table line based on the fully convolutional network (FCN). The problem of the former is that it is laborious to recognize the distorted table, while the problem of the latter is that the sample imbalance makes it difficult to train the model. In this paper, a gradient heatmap based table structure recognition method is proposed, by exploring the gradient heatmaps of the vertical lines and horizontal lines in the table. Specifically, the pixels of the vertical lines of the table are obtained according to the gradient heatmap, then the pixels of the horizontal lines are obtained using the same method, and finally the table structure is restored by using the connected domain search method. Compared with the Single Shot MultiBox Detector (SSD) and Faster RCNN that directly detects cells, our Average Precision (AP) value reached up to 99.5%, which is much higher than the above models. Additionally, we demonstrate that the AP values of the proposed models are reduced almost negligibly when the IoU threshold increased from 0.5 to 0.75, while the AP value of the fast RCNN and SSD model decreased significantly.