A lightweight table detection model based on Faster R-CNN for heterogeneous document images

Table detection refers to the process of identifying and localizing tables within documents, images, or other data sources. In this research, our focus is on table detection specifically tailored for heterogeneous document images. In the task of table detection, a typical requirement involves striking a balance between efficiency and accuracy. This research primarily focuses on the exploration of neural network architecture selection and optimization strategies for table detection, aiming to enhance efficiency while maintaining a high level of accuracy.. Firstly, we opt for the lightweight PP-LCNet network as the backbone for the table detection algorithm, thereby elevating the inference speed of the model when utilizing CPU resources. Subsequently, an optimization of the CornerNet architecture is undertaken, introducing a novel region proposal generation network denoted as CornerNet-lo. Following this, we employ the aforementioned networks to replace the region proposal network (RPN) in the Faster R-CNN framework. Consequently, our table detection approach attains state-of-the-art performance on the publicly available table detection benchmarks, PubLayNet and IIIT-AR-13K.