Table Detection in Complex Layouts

Table detection is a critical component of document automation that has garnered increasing attention. However, accurately detecting tables with arbitrary styles presents a significant challenge. To address the problem, we propose a method to detect tables in complex layouts based on the YOLOV5 model. We introduce Deformable Convolutional Networks (DCN) into the backbone of the network to enhance table feature extraction capability from tables with arbitrary layouts. Additionally, We design a new residual module, ResDCN, which replaces the residual cell structure in the original algorithm. Additionally, we employ a Global Attention Mechanism(GAM) to efficiently exploit semantic information from channel and spatial features. Lastly, We merge Adaptive Spatial Feature Fusion (ASFF) at the termination of the neck layer in the YOLOV5 model, promoting superior feature fusion across diverse scales. Our model shows exceptional performance in the ICDAR-2017POD table competition, with an F-measure(F1 score) of reaches 96.7%. Compared with the original YOLOV5 algorithm, our method improves the F1 score by 2.2% and attains a recall rate of 98% at a threshold of 0.6. 98% is state-of-the-art performance on the dataset. When evaluated at a higher threshold of 0.8, the F1 score reaches 96.4%, placing second in terms of the F1 score. Moreover, our model evaluates the TNCR dataset and achieves unparalleled performance, particularly in detecting complex wireless tables, where the F1 score improves by 2%. These noteworthy results show our model's proficiency in accurately detecting tables with diverse styles embedded in complex layouts.