search

qcf-568 / DocTamper

[CVPR2023] Towards Robust Tampered Text Detection in Document Image: New Dataset and New Solution
127 stars 11 forks source link

'rgb': np.clip(np.abs(dct),0,20) why？ #12

Closed 256-7421142 closed 11 months ago

Mohamed-Dhouib commented 1 year ago

Wondering the same thing !

qcf-568 commented 1 year ago
  1. Why use np.abs and np.clip: You can refer to the CAT-Netv2 paper. The most information of DCT coef focuses on about the zero point, and the sign of them are not so useful. 2. Why clip to 20: the same as CAT-Netv2 for a fair comparison. 3. Why named as 'rgb': It inherited from the debug process, it's actually the dct coef.