method: SPDB LAB2023-03-20
Authors: Jie Li 、Wei Wang、Yuqi Zhang、Ruixue Zhang、Yiru Zhao、Danya Zhou、Di Wang、Dong Xiang、Hui Wang、Min Xu、Pengyu Chen、Bin Zhang、Chao Li、Shiyu Hu、Songtao Li、Yunxin Yang
Affiliation: Shanghai Pudong Development Bank
Email: zhangyq26@outlook.com、wangdee0805@139.com、lij131@spdb.com.cn
Description: Circle seals and Ellipse seals：Based on the results of the circle and ellipse seals title detection in task1， PCA technology was used to correct the rotated seal, the image processing technology was used to separate the seal title, and finally the curved text was sent to the recognition model for recognition. The recognition model was selected by Trocr, and the training data includes the provided training data and synthetic data.
Rectangle seals and Triangle seals: rectangle seals and triangle seals were not based on the task1 detection model, but train a text line detection model. the image processing technology was used to separate the seal title. The recognition model was selected by Trocr, and the training data includes the provided by synthetic data.
method: rec320_32023-03-20
Authors: Huajian Zhou
Affiliation: China Mobile Cloud Centre
Description: detection model: vitdet(vit-huge)
recognition model: two recognize model (ctc , transformer)
direction model: use resnet for classification
text image rectify model: ellipse rectify algorithm
correct model: use country area info correct bad pred title
training data: official train_data, synthetic data
method: rec320_22023-03-20
Authors: Huajian Zhou
Affiliation: China Mobile Cloud Centre
Description: detection model: vitdet(vit-huge)
recognition model: two recognize model (ctc , transformer)
direction model: use resnet for classification
text image rectify model: ellipse rectify algorithm
correct model: use country area info correct bad pred title
training data: official train_data, synthetic data