Option --psm 10 digits are not taken account.

[ruimo]

Environment

• 4.0.0
• macOS 10.13.3

Current Behavior:

$ tesseract 000.png stdout --psm 10 digits
Warning: Invalid resolution 0 dpi. Using 70 instead.
iL

000.png:

Expected Behavior:

Since --psm 10 is specified, the output 'iL' is invalid. A single character should be recognized.
As 'digits' are specified as a config file, 'iL' is invalid. A digit character(0123456789-.) should be recognized.

Suggested Fix:

[amitdo]

The content of the digits config file:
tessedit_char_whitelist 0123456789-.

See #751.

[ruimo]

Thanks, I updated the expected behaviour.

[bertsky]

I have looked into this via debugging (adding DebugBeams at the end of RecodeBeamSearch::Decode). For the whitelisting, I used branch #2294.

Since --psm 10 is specified, the output 'iL' is invalid. A single character should be recognized.

True. From what I can tell, that PSM_SINGLE_CHAR does not work for LSTMs. And getting two characters out of one is a known error. I coined this the "diplopia" problem when presenting my analysis for #2339.

As 'digits' are specified as a config file, 'iL' is invalid. A digit character(0123456789-.) should be recognized.

True again. As @amitdo and @Shreeshrii pointed out, whitelisting will be re-instated for LSTMs after #2294 is merged. But even with that, for this example, nothing can be decoded (empty result), because the beam is too narrow.

It can be made wider, e.g. by changing ComputeTopN to not stop looking further than the first and second best input alternatives (which will be null and "." in this case). I do get a "1" as result with that modification, but the implications of this should be discussed in #2294.

[Shreeshrii]

@ruimo

Please try with my finetuned traineddata which is limited to digits and some punctuation (instead of digits config).

digits_layer.traineddata
digitsall_layer.traineddata

from https://github.com/Shreeshrii/tessdata_shreetest