Tesseract fine tuning error - Compute CTC targets failed - deep-learning

Question

I'm trying to fine-tune tesseract 4.1.1 on my own specific data according to this guide. I want it to become able to detect and recognize text in boxes like that:
I have generated a number of images like that and corresponding to them .box files containing bounding boxes with text. To reproduce my issue here i'm going to show my pipeline using only one image. Here is the .box file for the image above:
0 1804 1659 1858 1813 0
5 1804 1659 1858 1813 0
9 1804 1659 1858 1813 0
9 1266 715 1334 1169 0
7 1266 715 1334 1169 0
8 1266 715 1334 1169 0
3 1266 715 1334 1169 0
6 1266 715 1334 1169 0
8 1266 715 1334 1169 0
0 1266 715 1334 1169 0
5 1266 715 1334 1169 0
3 1266 715 1334 1169 0
2 876 303 930 607 0
7 876 303 930 607 0
2 876 303 930 607 0
8 876 303 930 607 0
2 876 303 930 607 0
2 876 303 930 607 0
8 1671 120 1725 224 0
0 1671 120 1725 224 0
5 300 1278 354 1482 0
2 300 1278 354 1482 0
3 300 1278 354 1482 0
7 300 1278 354 1482 0
7 917 1451 975 1605 0
6 917 1451 975 1605 0
4 917 1451 975 1605 0
1 1058 1310 1132 1716 0
9 1058 1310 1132 1716 0
8 1058 1310 1132 1716 0
7 1058 1310 1132 1716 0
7 1058 1310 1132 1716 0
1 1058 1310 1132 1716 0
8 1058 1310 1132 1716 0
6 1058 1310 1132 1716 0
3 998 76 1070 382 0
4 998 76 1070 382 0
4 998 76 1070 382 0
8 998 76 1070 382 0
3 998 76 1070 382 0
6 998 76 1070 382 0
3 722 548 776 652 0
2 722 548 776 652 0
7 1782 1332 1838 1586 0
7 1782 1332 1838 1586 0
2 1782 1332 1838 1586 0
6 1782 1332 1838 1586 0
2 1782 1332 1838 1586 0
1 714 140 768 244 0
2 714 140 768 244 0
0 220 500 278 754 0
5 220 500 278 754 0
5 220 500 278 754 0
6 220 500 278 754 0
6 220 500 278 754 0
8 1676 1052 1742 1406 0
4 1676 1052 1742 1406 0
5 1676 1052 1742 1406 0
9 1676 1052 1742 1406 0
1 1676 1052 1742 1406 0
2 1676 1052 1742 1406 0
4 1676 1052 1742 1406 0
5 357 161 419 317 0
1 357 161 419 317 0
4 357 161 419 317 0
9 1424 848 1480 952 0
8 1424 848 1480 952 0
0 438 324 498 478 0
6 438 324 498 478 0
9 438 324 498 478 0
8 1503 1246 1559 1700 0
1 1503 1246 1559 1700 0
8 1503 1246 1559 1700 0
5 1503 1246 1559 1700 0
3 1503 1246 1559 1700 0
0 1503 1246 1559 1700 0
5 1503 1246 1559 1700 0
5 1503 1246 1559 1700 0
4 1503 1246 1559 1700 0
8 1553 477 1609 581 0
4 1553 477 1609 581 0
3 527 258 581 512 0
7 527 258 581 512 0
7 527 258 581 512 0
9 527 258 581 512 0
1 527 258 581 512 0
6 1665 1592 1727 1748 0
8 1665 1592 1727 1748 0
3 1665 1592 1727 1748 0
5 595 1362 651 1766 0
9 595 1362 651 1766 0
3 595 1362 651 1766 0
9 595 1362 651 1766 0
4 595 1362 651 1766 0
3 595 1362 651 1766 0
3 595 1362 651 1766 0
1 595 1362 651 1766 0
I have also converted the image into .tiff format and placed it in the same directory with .box file. Lets say we have 87.tiff and 87.box inside the directory.
Next i generate 87.lstmf file using
tesseract 87.tiff 87 lstm.train
Next i extract model using
combine_tessdata -e /usr/share/tesseract-ocr/4.00/tessdata/rus.traineddata lstm_model/rus.lstm
Next i create train.txt file containing the single line: 87.lstmf
Finally, i create bash script train.sh
/usr/bin/lstmtraining \
--model_output output/fine_tuned \
--continue_from lstm_model/rus.lstm \
--traineddata /usr/share/tesseract-ocr/4.00/tessdata/rus.traineddata \
--train_listfile train.txt \
--eval_listfile train.txt \
--max_iterations 400\
--debug_level -1
And when i run it, i have the following logs:
$ bash train.sh
Loaded file lstm_model/rus.lstm, unpacking...
Warning: LSTMTrainer deserialized an LSTMRecognizer!
Continuing from lstm_model/rus.lstm
Loaded 1/1 lines (1-1) of document 87.lstmf
Loaded 1/1 lines (1-1) of document 87.lstmf
Compute CTC targets failed!
Compute CTC targets failed!
Compute CTC targets failed!
Compute CTC targets failed!
The message "Compute CTC targets failed!" repeats infinitely until i interrupt the script.
What am i doing wrong? I'm also concerned about message "Loaded 1/1 lines (1-1)" since i have multiple bounding boxes on the image.

Answer 1

For lstm training on images you may need to have the .box files in the new lstm format (these can be generated by running tesseract with the lstmbox argument):
TrainingTesseract-4.00
so after each line of text mark it with a special line:
<tab> <left> <bottom> <right> <top> <page>
then run lstm.train .