Table Structure Recognition Method Based on Lightweight Network and Channel Attention

The table recognition model rows and columns aggregated network (RCANet) uses a semantic segmentation approach to recognize structure, achieves better performance in row column segmentation. However, this ResNet18 as the backbone network, has 11.35 million parameters volume of 45.5 M, which is inconvenient deploy lightweight servers or mobile terminals. Therefore, from perspective compression, paper proposes attention (LRCAANet), ShuffleNetv2 replace original RCANet simplify size. Considering that reduces number feature channels, it certain impact on model. In order strengthen learning between (RAA) module (CAA) are proposed. RAA CAA add squeeze excitation (SE) modules, respectively. Adding SE means can learn correlation channels improve prediction effect experimental results show our method greatly while ensuring low-performance loss. end, average F1 score only 1.77% lower than model, 0.17 million, 0.8 M. Compared with parameter amount reduced by more 95%.