all repos — videocr @ 51ab92cca4e15de4695cb07c756488c6cd22f663

Extract hardcoded subtitles from videos using machine learning

videocr/video.py (view raw)

  1from __future__ import annotations
  2import multiprocessing
  3import pytesseract
  4import cv2
  5
  6from . import constants
  7from . import utils
  8from .models import PredictedFrame, PredictedSubtitle
  9from .opencv_adapter import Capture
 10
 11
 12class Video:
 13    path: str
 14    lang: str
 15    use_fullframe: bool
 16    num_frames: int
 17    fps: float
 18    height: int
 19    pred_frames: List[PredictedFrame]
 20    pred_subs: List[PredictedSubtitle]
 21
 22    def __init__(self, path: str):
 23        self.path = path
 24        with Capture(path) as v:
 25            self.num_frames = int(v.get(cv2.CAP_PROP_FRAME_COUNT))
 26            self.fps = v.get(cv2.CAP_PROP_FPS)
 27            self.height = int(v.get(cv2.CAP_PROP_FRAME_HEIGHT))
 28
 29    def run_ocr(self, lang: str, time_start: str, time_end: str,
 30                conf_threshold: int, use_fullframe: bool) -> None:
 31        self.lang = lang
 32        self.use_fullframe = use_fullframe
 33
 34        ocr_start = utils.get_frame_index(time_start, self.fps) if time_start else 0
 35        ocr_end = utils.get_frame_index(time_end, self.fps) if time_end else self.num_frames
 36
 37        if ocr_end < ocr_start:
 38            raise ValueError('time_start is later than time_end')
 39        num_ocr_frames = ocr_end - ocr_start
 40
 41        # get frames from ocr_start to ocr_end
 42        with Capture(self.path) as v, multiprocessing.Pool() as pool:
 43            v.set(cv2.CAP_PROP_POS_FRAMES, ocr_start)
 44            frames = (v.read()[1] for _ in range(num_ocr_frames))
 45
 46            # perform ocr to frames in parallel
 47            it_ocr = pool.imap(self._image_to_data, frames, chunksize=10)
 48            self.pred_frames = [
 49                PredictedFrame(i + ocr_start, data, conf_threshold)
 50                for i, data in enumerate(it_ocr)
 51            ]
 52
 53    def _image_to_data(self, img) -> str:
 54        if not self.use_fullframe:
 55            # only use bottom half of the frame by default
 56            img = img[self.height // 2:, :]
 57        config = '--tessdata-dir "{}"'.format(constants.TESSDATA_DIR)
 58        return pytesseract.image_to_data(img, lang=self.lang, config=config)
 59
 60    def get_subtitles(self, sim_threshold: int) -> str:
 61        self._generate_subtitles(sim_threshold)
 62        return ''.join(
 63            '{}\n{} --> {}\n{}\n\n'.format(
 64                i,
 65                utils.get_srt_timestamp(sub.index_start, self.fps),
 66                utils.get_srt_timestamp(sub.index_end, self.fps),
 67                sub.text)
 68            for i, sub in enumerate(self.pred_subs))
 69
 70    def _generate_subtitles(self, sim_threshold: int) -> None:
 71        self.pred_subs = []
 72
 73        if self.pred_frames is None:
 74            raise AttributeError(
 75                'Please call self.run_ocr() first to perform ocr on frames')
 76
 77        # divide ocr of frames into subtitle paragraphs using sliding window
 78        WIN_BOUND = int(self.fps // 2)  # 1/2 sec sliding window boundary
 79        bound = WIN_BOUND
 80        i = 0
 81        j = 1
 82        while j < len(self.pred_frames):
 83            fi, fj = self.pred_frames[i], self.pred_frames[j]
 84
 85            if fi.is_similar_to(fj):
 86                bound = WIN_BOUND
 87            elif bound > 0:
 88                bound -= 1
 89            else:
 90                # divide subtitle paragraphs
 91                para_new = j - WIN_BOUND
 92                self._append_sub(PredictedSubtitle(
 93                    self.pred_frames[i:para_new], sim_threshold))
 94                i = para_new
 95                j = i
 96                bound = WIN_BOUND
 97
 98            j += 1
 99
100        # also handle the last remaining frames
101        if i < len(self.pred_frames) - 1:
102            self._append_sub(PredictedSubtitle(
103                self.pred_frames[i:], sim_threshold))
104
105    def _append_sub(self, sub: PredictedSubtitle) -> None:
106        if len(sub.text) == 0:
107            return
108
109        # merge new sub to the last subs if they are similar
110        while self.pred_subs and sub.is_similar_to(self.pred_subs[-1]):
111            ls = self.pred_subs[-1]
112            del self.pred_subs[-1]
113            sub = PredictedSubtitle(ls.frames + sub.frames, sub.sim_threshold)
114
115        self.pred_subs.append(sub)