Authors: chen duan,pei fu,shan guo

Description: The detection model uses the DB framework and adds Transformer to further enhance feature aggregation ability. In addition, center point supervision and four corner point supervision are added. The model is trained on the SynthText dataset for 50 epochs, and then trained on public datasets such as ICDAR15, Total-Text, CTW1500, and MLT19 for 200 epochs. Finally, fine-tuning is performed for 200 epochs. In the training process, the AdamW optimizer is used, and the learning rate decay strategy is set. To further optimize the model's performance, some techniques are used, such as data augmentation