Authors: Jun Yu Xinjian Gao Yaohui Zhang Ye Pang

Affiliation: University of Science and Technology of China；PingAn Technology

Description: 1. We build fully Transformer Networks consisting a Vision Transformer encoder and Transformer decoder，which gives a better performance on curved seal text due to Transformer's powerful global self-attention mechanism.
2. We use a large number of data enhancement strategies, such as random rotation, random blur, brightness adjustment, etc., to simulate the rotation, bending and background text interference in the original training set.
3. We have made a large number of synthetic data sets to make up for the lack of training sets to improve the generalization ability of the model.
4. We use the classification network to classify the shapes of three different seals, and then use the recognition models to identify them respectively.