[Feature request] Efficient text detection

[Morv55555]

Efficient text detection

Currently the start and end of a subtitle is determined by the frames set by the user. Why not make the process more efficient by scanning frames at a higer rate (e.g. 30 fps video, check every 15th frame). Only when text detected, do a more thorough search around the frame to get the correct timestamps. This way the whole process should be way faster.

My suggestion would be to make a hybrid approach by combining coarse sampling pass with your SSIM filtering:

Coarse sampling pass

• From a 30fps video, pick every Nth frame (e.g. every 15th = 2 fps).
• Run OCR only on these frames.
• This gives a rough map of where subtitles appear/disappear.

Trigger expansion around detections

• If OCR finds text in a sampled frame:
• Look at its neighbors (e.g. ±10 frames).
• Run OCR on each of those frames to pinpoint the exact subtitle start/end timestamps with SSIM to avoid re-OCRing identical frames.

Merging step
Combine text blocks that remain the same across multiple consecutive frames into a single subtitle event.

[techguru0]

doing it twice will most likely end up taking the same amount of time

[Morv55555]

You are not doing it twice. You do only text detection first, no OCR, when you have the rough map, you do a more thorough search with OCR only on parts with subtitles. If you have a movie with (very) few subtitles, you basically can speed up and skip a lot e.g. if you have a 2h movie with only 300 subtitles.

[techguru0]

you are doing it twice -- the first check is still a full check

[Morv55555]

You could do both in one go but 2 is also possible and still faster. There is a reason vsf search + PaddleOCR is way faster than using VideOCR for both (and equal quality)