Optimizing Speed/Accuracy Table Detection via Knowledge Distillation

To solve the problem of insufficient real-time response to table detection model in edge devices, a lightweight method of table detection based on knowledge distillation was proposed in this paper, which can improve the response speed while ensuring the accuracy of the detection model. In addition to output table boundary simply, this paper added branch for the regression of table positioning to coordinate, so that the labels output by the teacher model can distill the student model at a label level. In addition to the label-level distillation, this paper also used the attention mechanism to implement the feature layer output of the teacher model, which can be learned by student model. At present, the style of table detection dataset is relatively single, and the resources of open dataset are insufficient. This paper proposed an open dataset to verify the lightweight model proposed in this paper. This paper conducted experiments on two open datasets, ICDAR 2017 and ICDAR 2019, as well as the proposed open dataset. F1 value was improved by an average of 1.7% on the three datasets, at the same time reached real-time speeds (all greater than 24FPS) on different devices. The experimental results show that the lightweight table detection model based on knowledge distillation can guarantee the accuracy of the model while ensuring the response speed.