Do you have char mask when the dataset is total-text?

[zty8599]

Hello,i use your model to train my own datastes , that include character-level segmentation annotations and word-level annotations for same data type. When i training datasets only include word-level annotations like total-text ,the recognition result is not good . But when two types of annotations are uesd,the result is better. Why? We guess the reason is char mask, when dataset is synthtext,char mask is created by annotations ; when dataset is total-text,char mask is created by neural network，the accuracy of char mask is low, so the final recognition result is bad. Right?Or when dataset is total-text,there is no char mask. Looking forward to your reply，thanks.