Improve Text Detection Model

[oggyfaker]

Hi @Topdu , thank for yours great repo, it helps me a lot.

• I want improve the model for my dataset because my dataset have a lot of dense text (text line very close and have multi oriented like horizontal - vertical - diagonal text). I saw that when i try training fine-tunning more epoch (100 - 200 epoch) on my dataset , the issue still exist, model always have 3 issue:

• With 2 lines text are close , they predict the shrink map mask in the gap of 2 text --> so that create the problem when model find the contours in post process

• With text have diagonal direction , they always fail in shrink map segmentation like this (missing or merge 2 lines text)

• Final issue , model always miss-detection with the single character like "A", "B", ... or single number like "1", "2". Maybe these character stand alone , it look like another object in the drawing dataset.

Here is what i want to ask:

• Do you think can i change the another backbone can improve the performance and fix all the issue above? (accept that speed can be lower). And if yes, what criteria points for choosing model ? ( feature maps must be larger than RepViT M0.9 you are using or something else? )
• For improve the broken case of shrink map (diagonal text), do you think what parameter i can experiment to update for DB loss to improve that , or DB model always meet this issue so i should find new model ?

Thanks in advance !