1from __future__ import annotations
  2from concurrent import futures
  3import datetime
  4import pytesseract
  5import cv2
  6
  7from . import constants
  8from .models import PredictedFrame, PredictedSubtitle
  9
 10
 11class Video:
 12    path: str
 13    lang: str
 14    use_fullframe: bool
 15    num_frames: int
 16    fps: float
 17    pred_frames: List[PredictedFrame]
 18    pred_subs: List[PredictedSubtitle]
 19
 20    def __init__(self, path: str):
 21        self.path = path
 22        v = cv2.VideoCapture(path)
 23        self.num_frames = int(v.get(cv2.CAP_PROP_FRAME_COUNT))
 24        self.fps = v.get(cv2.CAP_PROP_FPS)
 25        v.release()
 26
 27    def run_ocr(self, lang: str, time_start: str, time_end: str,
 28                use_fullframe: bool) -> None:
 29        self.lang = lang
 30        self.use_fullframe = use_fullframe
 31
 32        ocr_start = self._frame_index(time_start) if time_start else 0
 33        ocr_end = self._frame_index(time_end) if time_end else self.num_frames
 34
 35        if ocr_end < ocr_start:
 36            raise ValueError('time_start is later than time_end')
 37        num_ocr_frames = ocr_end - ocr_start
 38
 39        # get frames from ocr_start to ocr_end
 40        v = cv2.VideoCapture(self.path)
 41        v.set(cv2.CAP_PROP_POS_FRAMES, ocr_start)
 42        frames = (v.read()[1] for _ in range(num_ocr_frames))
 43
 44        # perform ocr to frames in parallel
 45        with futures.ProcessPoolExecutor() as pool:
 46            ocr_map = pool.map(self._single_frame_ocr, frames, chunksize=10)
 47            self.pred_frames = [PredictedFrame(i + ocr_start, data) 
 48                                for i, data in enumerate(ocr_map)]
 49
 50        v.release()
 51
 52    # convert time str to frame index
 53    def _frame_index(self, time: str) -> int:
 54        t = time.split(':')
 55        t = list(map(float, t))
 56        if len(t) == 3:
 57            td = datetime.timedelta(hours=t[0], minutes=t[1], seconds=t[2])
 58        elif len(t) == 2:
 59            td = datetime.timedelta(minutes=t[0], seconds=t[1])
 60        else:
 61            raise ValueError(
 62                'time data "{}" does not match format "%H:%M:%S"'.format(time))
 63
 64        index = int(td.total_seconds() * self.fps)
 65        if index > self.num_frames or index < 0:
 66            raise ValueError(
 67                'time data "{}" exceeds video duration'.format(time))
 68
 69        return index
 70
 71    def _single_frame_ocr(self, img) -> str:
 72        if not self.use_fullframe:
 73            # only use bottom half of the frame by default
 74            img = img[self.height // 2:, :]
 75        config = '--tessdata-dir "{}"'.format(constants.TESSDATA_DIR)
 76        return pytesseract.image_to_data(img, lang=self.lang, config=config)
 77
 78    def get_subtitles(self) -> str:
 79        self._generate_subtitles()
 80        return ''.join(
 81            '{}\n{} --> {}\n{}\n\n'.format(
 82                i,
 83                self._srt_timestamp(sub.index_start),
 84                self._srt_timestamp(sub.index_end),
 85                sub.text)
 86            for i, sub in enumerate(self.pred_subs))
 87
 88    def _generate_subtitles(self) -> None:
 89        self.pred_subs = []
 90
 91        if self.pred_frames is None:
 92            raise AttributeError(
 93                'Please call self.run_ocr() first to perform ocr on frames')
 94
 95        # divide ocr of frames into subtitle paragraphs using sliding window
 96        WIN_BOUND = int(self.fps // 2)  # 1/2 sec sliding window boundary
 97        bound = WIN_BOUND
 98        i = 0
 99        j = 1
100        while j < len(self.pred_frames):
101            fi, fj = self.pred_frames[i], self.pred_frames[j]
102
103            if fi.is_similar_to(fj):
104                bound = WIN_BOUND
105            elif bound > 0:
106                bound -= 1
107            else:
108                # divide subtitle paragraphs
109                para_new = j - WIN_BOUND
110                self._append_sub(
111                    PredictedSubtitle(self.pred_frames[i:para_new]))
112                i = para_new
113                j = i
114                bound = WIN_BOUND
115
116            j += 1
117
118        # also handle the last remaining frames
119        if i < len(self.pred_frames) - 1:
120            self._append_sub(PredictedSubtitle(self.pred_frames[i:]))
121
122    def _append_sub(self, sub: PredictedSubtitle) -> None:
123        if len(sub.text) == 0:
124            return
125
126        # merge new sub to the last subs if they are similar
127        while self.pred_subs and sub.is_similar_to(self.pred_subs[-1]):
128            ls = self.pred_subs[-1]
129            del self.pred_subs[-1]
130            sub = PredictedSubtitle(ls.frames + sub.frames)
131
132        self.pred_subs.append(sub)
133
134    def _srt_timestamp(self, frame_index: int) -> str:
135        td = datetime.timedelta(seconds=frame_index / self.fps)
136        ms = td.microseconds // 1000
137        m, s = divmod(td.seconds, 60)
138        h, m = divmod(m, 60)
139        return '{:02d}:{:02d}:{:02d},{:03d}'.format(h, m, s, ms)