Box_padding is shift on the fisrt row/column

[zwwlouis]

Tesseract Open Source OCR Engine v4.0.0-beta.4-18-g4370
text2image 4.0.0-beta.4-18-g4370

Current Behavior:

Since I'v found that model is sensitive on img padding, I'm trying to add some random box_padding when rendering img
But the words on the first row or column are wrong in padding, as the picture below.

Some words are always on the first line, So they are always geting wrong box info which will likely lead to overfiting as I know.

Expected Behavior:

I expect box_padding to be uniformly distribute no matter the word is on the edge or inside

[amitdo]

The bounding boxes of characters are not very reliable when the lstm engine is being used for text recognition.

Do you get the right text output?

[zwwlouis]

Nope, I use eng training text to make an example. For chinese word it always box up character by character.
This box shift influence my model more or less.

The bounding boxes of characters are not very reliable when the lstm engine is being used for text
recognition.

Do you get the right text
output?