Table Detection Method Based on Faster-RCNN and Window Attention

As an important carrier of information, tables possess the characteristics of high data storage density, conciseness, and intuitiveness, and are widely applied in offices and daily life. Due to the complexity of table structures and diverse presentation formats, the automated processing of a large number of image-based tables has always been a challenge in the field of document recognition. This algorithm addresses the task of table detection in table processing and proposes a table detection algorithm based on an improved window self-attention network for feature extraction of image-based tables. It utilizes a two-stage object detection algorithm, introduces local feature extraction blocks and backward feed-forward residual network blocks, and designs a feature pyramid network within the backbone to enhance the model's detection performance by improving its ability to learn document spatial layout features. The effectiveness of the proposed method is verified through experimental comparisons on publicly available datasets.