Description
Hello,
First of all, thank you for the continuous work on Subtitle Edit.
While testing the new Paddle OCR Standalone engine with Japanese Blu-ray PGS subtitles, I noticed a reproducible behavior related to punctuation recognition.
Overall, Paddle OCR produces significantly better Japanese text recognition than Tesseract OCR and another OCR solution I tested. However, punctuation located at the beginning or end of subtitle lines is frequently omitted.
Environment
Subtitle Edit 5.1.0-rc4
Subtitle Edit 5.1.0-rc7 (same behavior)
OCR engine: Paddle OCR Standalone
Language: Japanese
Source: Japanese Blu-ray PGS (.sup)
Operating System: Linux (Linux Lite)
Observed behavior
The following punctuation is frequently omitted when it appears near subtitle line boundaries (most commonly at the end of a subtitle line):
Ellipsis (... / …)
Leading or trailing horizontal dashes (ー / dialogue dash)
Examples:
Example 1
Image:
(ジンコ) はしょったね...
OCR result:
(ジンコ) はしょったね
The ellipsis is omitted.
Example 2
Image:
あー あのさ...
OCR result:
あー あのさ
The trailing ellipsis is omitted.
Example 3
Image:
(店長) あのー
OCR result:
(店長) あの
The trailing horizontal dash is omitted.
Interesting observation
The behavior appears to depend on the punctuation's position within the recognized text rather than on the character itself, since the same punctuation is recognized correctly in other contexts.
For example, ellipsis characters located inside a sentence are recognized correctly, whereas the same character is often omitted when it appears at the end of a subtitle line.
The same punctuation is often recognized correctly when it is surrounded by other characters.
For example:
Image:
なんか…恥ずかしいんだよね
OCR:
なんか…恥ずかしいんだよね
The ellipsis is preserved.
Additional observations
The behavior is not completely deterministic.
For two-line subtitles:
sometimes the ellipsis at the end of the second line is omitted;
in other cases it is recognized correctly.
Therefore, the issue does not seem to depend only on line length or subtitle layout.
What has already been tested
Same result in rc4 and rc7.
Image preprocessing options (binarization, cropping, etc.) do not change the behavior.
The punctuation is clearly visible in the subtitle image.
For comparison, I also tested Tesseract OCR and another OCR solution on the same PGS subtitles. They correctly recognize these punctuation marks and do not appear to exhibit this specific behavior, suggesting that the subtitle images themselves are not the cause.
Question
Could this behavior be related to Paddle OCR post-processing, or could Subtitle Edit be trimming punctuation located at subtitle line boundaries?
If this is expected Paddle OCR behavior, is there any configuration or parameter that could preserve this punctuation?
If needed, I can provide the original PGS subtitle together with additional screenshots and OCR output.
I can also provide a minimal sample containing only the affected subtitle events if that would make testing easier.
Thank you!
Description
Hello,
First of all, thank you for the continuous work on Subtitle Edit.
While testing the new Paddle OCR Standalone engine with Japanese Blu-ray PGS subtitles, I noticed a reproducible behavior related to punctuation recognition.
Overall, Paddle OCR produces significantly better Japanese text recognition than Tesseract OCR and another OCR solution I tested. However, punctuation located at the beginning or end of subtitle lines is frequently omitted.
Environment
Subtitle Edit 5.1.0-rc4
Subtitle Edit 5.1.0-rc7 (same behavior)
OCR engine: Paddle OCR Standalone
Language: Japanese
Source: Japanese Blu-ray PGS (.sup)
Operating System: Linux (Linux Lite)
Observed behavior
The following punctuation is frequently omitted when it appears near subtitle line boundaries (most commonly at the end of a subtitle line):
Ellipsis (... / …)
Leading or trailing horizontal dashes (ー / dialogue dash)
Examples:
Example 1
Image:
(ジンコ) はしょったね...
OCR result:
(ジンコ) はしょったね
The ellipsis is omitted.
Example 2
Image:
あー あのさ...
OCR result:
あー あのさ
The trailing ellipsis is omitted.
Example 3
Image:
(店長) あのー
OCR result:
(店長) あの
The trailing horizontal dash is omitted.
Interesting observation
The behavior appears to depend on the punctuation's position within the recognized text rather than on the character itself, since the same punctuation is recognized correctly in other contexts.
For example, ellipsis characters located inside a sentence are recognized correctly, whereas the same character is often omitted when it appears at the end of a subtitle line.
The same punctuation is often recognized correctly when it is surrounded by other characters.
For example:
Image:
なんか…恥ずかしいんだよね
OCR:
なんか…恥ずかしいんだよね
The ellipsis is preserved.
Additional observations
The behavior is not completely deterministic.
For two-line subtitles:
sometimes the ellipsis at the end of the second line is omitted;
in other cases it is recognized correctly.
Therefore, the issue does not seem to depend only on line length or subtitle layout.
What has already been tested
Same result in rc4 and rc7.
Image preprocessing options (binarization, cropping, etc.) do not change the behavior.
The punctuation is clearly visible in the subtitle image.
For comparison, I also tested Tesseract OCR and another OCR solution on the same PGS subtitles. They correctly recognize these punctuation marks and do not appear to exhibit this specific behavior, suggesting that the subtitle images themselves are not the cause.
Question
Could this behavior be related to Paddle OCR post-processing, or could Subtitle Edit be trimming punctuation located at subtitle line boundaries?
If this is expected Paddle OCR behavior, is there any configuration or parameter that could preserve this punctuation?
If needed, I can provide the original PGS subtitle together with additional screenshots and OCR output.
I can also provide a minimal sample containing only the affected subtitle events if that would make testing easier.
Thank you!