Investigate if continuous slice of frames with the same subtitles can be detected faster using cv solution

[devmaxxing]

If so, then PaddleOCR only needs to be run on one frame per subtitle.

[SWHL]

• I have the same idea. So I create the RapidVideOCR repo.
• Welcome to study together.

[devmaxxing]

@SWHL Sure. I'm trying to focus on another project at the moment but will definitely come back to this later. In the meantime, I guess some vague initial ideas I had on this was:

• we could loop through each frame, checking against and adding to a single "accumulator" frame for each iteration. If we detect a subtitle change, then we process the accumulator frame and replace it with the current frame. If no subtitle change is detected, then we incorporate the current frame into the accumulator frame
• background pixels from previous frames could be useful for removing noise on the current frame
• it's ok if there are occasional false positives (identifying a frame as a subtitle change when it is not), just need to have some post processing that takes them into consideration like merging duplicate consecutive frame subtitles and subtitles with durations that are too short
• maybe there's some existing research papers that solves similar problems (isolation of static shapes in a video feed) out there