search

tesseract-ocr / tesstrain

Train Tesseract LSTM with make
Apache License 2.0
625 stars 180 forks source link

Traineddata for a dot matrix font taken with a Raspicam #184

Closed mbretschn closed 3 years ago

mbretschn commented 4 years ago

I try to build a traineddata model for a dot matrix font with custom box data. I had no success building traineddata which can recognize a training image shown below. I wonder, if it is even possible to create traineddata for the given images.

The pictures only contain numbers and were taken with a Raspicam. I set the value for DPI to 600. The complete ground-truth data can be viewed here.

Am example for a single image

imagedata

Boxdata

0 37 26 132 159 0
9 166 29 261 162 0

.gt.txt (i tried also without a blank between the numbers)

0 9

In the following the output of make training

> make training 
find data/sabei-ground-truth -name '*.gt.txt' | xargs cat | sort | uniq > "data/sabei/all-gt"
unicharset_extractor --output_unicharset "data/sabei/unicharset" --norm_mode 2 "data/sabei/all-gt"
Bad box coordinates in boxfile string! 23171014343530070829312213062132042527261519002033030116241109021205283618
Extracting unicharset from plain text file data/sabei/all-gt
Wrote unicharset file data/sabei/unicharset
+ tesseract data/sabei-ground-truth/23.png data/sabei-ground-truth/23 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/17.png data/sabei-ground-truth/17 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/10.png data/sabei-ground-truth/10 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/14.png data/sabei-ground-truth/14 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/34.png data/sabei-ground-truth/34 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/35.png data/sabei-ground-truth/35 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/30.png data/sabei-ground-truth/30 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/07.png data/sabei-ground-truth/07 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/08.png data/sabei-ground-truth/08 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/29.png data/sabei-ground-truth/29 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/31.png data/sabei-ground-truth/31 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/22.png data/sabei-ground-truth/22 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/13.png data/sabei-ground-truth/13 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/06.png data/sabei-ground-truth/06 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/21.png data/sabei-ground-truth/21 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/32.png data/sabei-ground-truth/32 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/04.png data/sabei-ground-truth/04 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/25.png data/sabei-ground-truth/25 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/27.png data/sabei-ground-truth/27 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/26.png data/sabei-ground-truth/26 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/15.png data/sabei-ground-truth/15 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/19.png data/sabei-ground-truth/19 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/00.png data/sabei-ground-truth/00 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/20.png data/sabei-ground-truth/20 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/33.png data/sabei-ground-truth/33 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/03.png data/sabei-ground-truth/03 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/01.png data/sabei-ground-truth/01 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/16.png data/sabei-ground-truth/16 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/24.png data/sabei-ground-truth/24 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/11.png data/sabei-ground-truth/11 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/09.png data/sabei-ground-truth/09 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/02.png data/sabei-ground-truth/02 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/12.png data/sabei-ground-truth/12 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/05.png data/sabei-ground-truth/05 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/28.png data/sabei-ground-truth/28 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/36.png data/sabei-ground-truth/36 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
+ tesseract data/sabei-ground-truth/18.png data/sabei-ground-truth/18 --psm 13 lstm.train
Tesseract Open Source OCR Engine v4.1.1 with Leptonica
find data/sabei-ground-truth -name '*.lstmf' | python3 shuffle.py 0 > "data/sabei/all-lstmf"
+ head -n 33 data/sabei/all-lstmf
+ tail -n 4 data/sabei/all-lstmf
wget -Odata/radical-stroke.txt 'https://github.com/tesseract-ocr/langdata_lstm/raw/master/radical-stroke.txt'
--2020-08-14 13:33:13--  https://github.com/tesseract-ocr/langdata_lstm/raw/master/radical-stroke.txt
Resolving github.com (github.com)... 140.82.118.3
Connecting to github.com (github.com)|140.82.118.3|:443... connected.
HTTP request sent, awaiting response... 302 Found
Location: https://raw.githubusercontent.com/tesseract-ocr/langdata_lstm/master/radical-stroke.txt [following]
--2020-08-14 13:33:13--  https://raw.githubusercontent.com/tesseract-ocr/langdata_lstm/master/radical-stroke.txt
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 151.101.112.133
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|151.101.112.133|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 330874 (323K) [text/plain]
Saving to: 'data/radical-stroke.txt'

     0K .......... .......... .......... .......... .......... 15% 1.24M 0s
    50K .......... .......... .......... .......... .......... 30% 1.40M 0s
   100K .......... .......... .......... .......... .......... 46% 2.05M 0s
   150K .......... .......... .......... .......... .......... 61% 2.60M 0s
   200K .......... .......... .......... .......... .......... 77% 4.68M 0s
   250K .......... .......... .......... .......... .......... 92% 1.54M 0s
   300K .......... .......... ...                             100% 1.74M=0.2s

2020-08-14 13:33:14 (1.84 MB/s) - 'data/radical-stroke.txt' saved [330874/330874]

combine_lang_model \
  --input_unicharset data/sabei/unicharset \
  --script_dir data \
  --numbers data/sabei/sabei.numbers \
  --puncs data/sabei/sabei.punc \
  --words data/sabei/sabei.wordlist \
  --output_dir data \
   \
  --lang sabei
Failed to read data from: data/sabei/sabei.wordlist
Failed to read data from: data/sabei/sabei.punc
Failed to read data from: data/sabei/sabei.numbers
Loaded unicharset of size 13 from file data/sabei/unicharset
Setting unichar properties
Setting script properties
Failed to load script unicharset from:data/Latin.unicharset
Warning: properties incomplete for index 3 = 2
Warning: properties incomplete for index 4 = 3
Warning: properties incomplete for index 5 = 1
Warning: properties incomplete for index 6 = 7
Warning: properties incomplete for index 7 = 0
Warning: properties incomplete for index 8 = 4
Warning: properties incomplete for index 9 = 5
Warning: properties incomplete for index 10 = 8
Warning: properties incomplete for index 11 = 9
Warning: properties incomplete for index 12 = 6
Config file is optional, continuing...
Failed to read data from: data/sabei/sabei.config
Null char=2
lstmtraining \
  --debug_interval 0 \
  --traineddata data/sabei/sabei.traineddata \
  --learning_rate 0.002 \
  --net_spec "[1,36,0,1 Ct3,3,16 Mp3,3 Lfys48 Lfx96 Lrx96 Lfx256 O1c`head -n1 data/sabei/unicharset`]" \
  --model_output data/sabei/checkpoints/sabei \
  --train_listfile data/sabei/list.train \
  --eval_listfile data/sabei/list.eval \
  --max_iterations 10000 \
  --target_error_rate 0.01
Warning: given outputs 13 not equal to unicharset of 12.
Num outputs,weights in Series:
  1,36,0,1:1, 0
Num outputs,weights in Series:
  C3,3:9, 0
  Ft16:16, 160
Total weights = 160
  [C3,3Ft16]:16, 160
  Mp3,3:16, 0
  Lfys48:48, 12480
  Lfx96:96, 55680
  Lrx96:96, 74112
  Lfx256:256, 361472
  Fc12:12, 3084
Total weights = 506988
Built network:[1,36,0,1[C3,3Ft16]Mp3,3Lfys48Lfx96Lrx96Lfx256Fc12] from request [1,36,0,1 Ct3,3,16 Mp3,3 Lfys48 Lfx96 Lrx96 Lfx256 O1c13]
Training parameters:
  Debug interval = 0, weights = 0.1, learning rate = 0.002, momentum=0.5
null char=11
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/30.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/25.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/17.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/15.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/18.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/02.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/21.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/09.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/28.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/36.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/16.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/20.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/31.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/35.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/19.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/14.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/22.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/34.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/24.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/10.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/33.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/29.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/11.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/32.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/01.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/08.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/06.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/12.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/26.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/07.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/00.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/27.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/03.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/04.lstmf
Loaded 1/1 lines (1-1) of document data/sabei-ground-truth/13.lstmf
At iteration 14/100/100, Mean rms=11.122%, delta=1.842%, char train=120.5%, word train=100%, skip ratio=0%,  New worst char error = 120.5 wrote checkpoint.

At iteration 88/200/200, Mean rms=10.953%, delta=4.947%, char train=118.25%, word train=100%, skip ratio=0%,  New worst char error = 118.25 wrote checkpoint.

At iteration 173/300/300, Mean rms=10.82%, delta=6.842%, char train=116.667%, word train=99.333%, skip ratio=0%,  New worst char error = 116.667 wrote checkpoint.

At iteration 267/400/400, Mean rms=10.482%, delta=8.737%, char train=113.25%, word train=95.5%, skip ratio=0%,  New worst char error = 113.25 wrote checkpoint.

At iteration 352/500/500, Mean rms=9.935%, delta=8.895%, char train=107.7%, word train=92.4%, skip ratio=0%,  New worst char error = 107.7 wrote checkpoint.

At iteration 427/600/600, Mean rms=9.388%, delta=8.825%, char train=101.167%, word train=88.333%, skip ratio=0%,  New worst char error = 101.167 wrote checkpoint.

2 Percent improvement time=501, best error was 100 @ 0
At iteration 501/700/700, Mean rms=9.012%, delta=8.857%, char train=96.714%, word train=85%, skip ratio=0%,  New best char error = 96.714 wrote checkpoint.

2 Percent improvement time=58, best error was 96.714 @ 501
At iteration 559/800/800, Mean rms=8.519%, delta=8.329%, char train=90.5%, word train=80%, skip ratio=0%,  New best char error = 90.5 wrote checkpoint.

2 Percent improvement time=59, best error was 90.5 @ 559
At iteration 618/900/900, Mean rms=8.131%, delta=8.094%, char train=85.889%, word train=76.333%, skip ratio=0%,  New best char error = 85.889 wrote checkpoint.

2 Percent improvement time=67, best error was 85.889 @ 618
At iteration 685/1000/1000, Mean rms=7.894%, delta=7.947%, char train=82.15%, word train=73.8%, skip ratio=0%,  New best char error = 82.15 wrote checkpoint.

2 Percent improvement time=49, best error was 82.15 @ 685
At iteration 734/1100/1100, Mean rms=7.236%, delta=8.305%, char train=73.5%, word train=67.2%, skip ratio=0%,  New best char error = 73.5 wrote checkpoint.

2 Percent improvement time=37, best error was 73.5 @ 734
At iteration 771/1200/1200, Mean rms=6.504%, delta=7.858%, char train=64.35%, word train=59.7%, skip ratio=0%,  New best char error = 64.35 wrote best model:data/sabei/checkpoints/sabei64.35_771.checkpoint wrote checkpoint.

2 Percent improvement time=23, best error was 64.35 @ 771
At iteration 794/1300/1300, Mean rms=5.71%, delta=7.037%, char train=54.55%, word train=51.5%, skip ratio=0%,  New best char error = 54.55 wrote best model:data/sabei/checkpoints/sabei54.55_794.checkpoint wrote checkpoint.

2 Percent improvement time=19, best error was 54.55 @ 794
At iteration 813/1400/1400, Mean rms=5.021%, delta=5.874%, char train=45.75%, word train=44.7%, skip ratio=0%,  New best char error = 45.75 wrote best model:data/sabei/checkpoints/sabei45.75_813.checkpoint wrote checkpoint.

2 Percent improvement time=35, best error was 45.75 @ 813
At iteration 848/1500/1500, Mean rms=4.589%, delta=5.258%, char train=40.05%, word train=39.5%, skip ratio=0%,  New best char error = 40.05 wrote best model:data/sabei/checkpoints/sabei40.05_848.checkpoint wrote checkpoint.

2 Percent improvement time=9, best error was 40.05 @ 848
At iteration 857/1600/1600, Mean rms=4.105%, delta=4.495%, char train=33.6%, word train=33.1%, skip ratio=0%,  New best char error = 33.6 wrote best model:data/sabei/checkpoints/sabei33.6_857.checkpoint wrote checkpoint.

2 Percent improvement time=3, best error was 33.6 @ 857
At iteration 860/1700/1700, Mean rms=3.537%, delta=3.626%, char train=26.8%, word train=26.8%, skip ratio=0%,  New best char error = 26.8 wrote best model:data/sabei/checkpoints/sabei26.8_860.checkpoint wrote checkpoint.

2 Percent improvement time=0, best error was 26.8 @ 860
At iteration 860/1800/1800, Mean rms=3.07%, delta=3.163%, char train=22.1%, word train=22.3%, skip ratio=0%,  New best char error = 22.1 wrote best model:data/sabei/checkpoints/sabei22.1_860.checkpoint wrote checkpoint.

2 Percent improvement time=0, best error was 22.1 @ 860
At iteration 860/1900/1900, Mean rms=2.593%, delta=2.542%, char train=17.2%, word train=17.6%, skip ratio=0%,  New best char error = 17.2 wrote best model:data/sabei/checkpoints/sabei17.2_860.checkpoint wrote checkpoint.

2 Percent improvement time=0, best error was 17.2 @ 860
At iteration 860/2000/2000, Mean rms=2.039%, delta=1.879%, char train=12.35%, word train=12.5%, skip ratio=0%,  New best char error = 12.35 wrote best model:data/sabei/checkpoints/sabei12.35_860.checkpoint wrote checkpoint.

2 Percent improvement time=0, best error was 12.35 @ 860
At iteration 860/2100/2100, Mean rms=1.603%, delta=1.337%, char train=8.95%, word train=9.1%, skip ratio=0%,  New best char error = 8.95 Transitioned to stage 1 wrote best model:data/sabei/checkpoints/sabei8.95_860.checkpoint wrote checkpoint.

2 Percent improvement time=0, best error was 8.95 @ 860
At iteration 860/2200/2200, Mean rms=1.275%, delta=0.979%, char train=6.5%, word train=6.6%, skip ratio=0%,  New best char error = 6.5 wrote best model:data/sabei/checkpoints/sabei6.5_860.checkpoint wrote checkpoint.

2 Percent improvement time=0, best error was 8.95 @ 860
At iteration 860/2300/2300, Mean rms=1.029%, delta=0.737%, char train=4.95%, word train=5%, skip ratio=0%,  New best char error = 4.95 wrote best model:data/sabei/checkpoints/sabei4.95_860.checkpoint wrote checkpoint.

2 Percent improvement time=0, best error was 6.5 @ 860
At iteration 860/2400/2400, Mean rms=0.786%, delta=0.458%, char train=3.45%, word train=3.4%, skip ratio=0%,  New best char error = 3.45 wrote best model:data/sabei/checkpoints/sabei3.45_860.checkpoint wrote checkpoint.

2 Percent improvement time=0, best error was 3.45 @ 860
At iteration 860/2500/2500, Mean rms=0.459%, delta=0.121%, char train=0.6%, word train=0.6%, skip ratio=0%,  New best char error = 0.6 wrote best model:data/sabei/checkpoints/sabei0.6_860.checkpoint wrote checkpoint.

2 Percent improvement time=0, best error was 3.45 @ 860
At iteration 860/2600/2600, Mean rms=0.292%, delta=0.037%, char train=0.2%, word train=0.2%, skip ratio=0%,  New best char error = 0.2 wrote best model:data/sabei/checkpoints/sabei0.2_860.checkpoint wrote checkpoint.

2 Percent improvement time=0, best error was 3.45 @ 860
At iteration 860/2700/2700, Mean rms=0.198%, delta=0%, char train=0%, word train=0%, skip ratio=0%,  New best char error = 0 wrote best model:data/sabei/checkpoints/sabei0_860.checkpoint wrote checkpoint.

Finished! Error rate = 0
lstmtraining \
--stop_training \
--continue_from data/sabei/checkpoints/sabei_checkpoint \
--traineddata data/sabei/sabei.traineddata \
--model_output data/sabei.traineddata
Loaded file data/sabei/checkpoints/sabei_checkpoint, unpacking...
`

The ground-truth files can be reviewed [`here`](https://github.com/3epnm/sabei)
wrznr commented 4 years ago

I had no success building traineddata which can recognize a training image shown below.

What does that mean? (From what I see in your output of the training, everything looks as expected.)

Pls. also note that your box file does meet the expectations of tesstrain which is a tool for training lines of text. At least the closing line indicating EOF is missing. Did you build your boxfile with tesstrain?

mbretschn commented 4 years ago

Did you build your boxfile with tesstrain?

No, The box data is generated.

stale[bot] commented 3 years ago

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.