I have trained a SATRN_small model on a custom dataset using MMOCR framework for character recognition as well as a SVRT_LCNet model using PaddleOCR.
The loss function used in MMOCR is CrossEntropyLoss, and the loss function used by PaddleOCR is a combination of CTC loss and SAR loss.
The repeating characters in a word are being suppressed by both models. for e.g, 40799932 is predicted as 4079932.
What is the reason for this problem? Is there any solution?