all repos — videocr @ a5e6845a1bc3d3497bda2e87b87bbeaf2541b3f2

Extract hardcoded subtitles from videos using machine learning

videocr/models.py (view raw)

 1from __future__ import annotations
 2from typing import List
 3from dataclasses import dataclass
 4from fuzzywuzzy import fuzz
 5
 6
 7@dataclass
 8class PredictedWord:
 9    __slots__ = 'confidence', 'text'
10    confidence: int
11    text: str
12
13
14class PredictedFrame:
15    index: int  # 0-based index of the frame
16    words: List[PredictedWord]
17    confidence: int  # total confidence of all words
18    text: str
19
20    def __init__(self, index: int, pred_data: str, conf_threshold=70):
21        self.index = index
22        self.words = []
23
24        block = 0  # keep track of line breaks
25
26        for l in pred_data.splitlines()[1:]:
27            word_data = l.split()
28            if len(word_data) < 12:
29                # no word is predicted
30                continue
31            _, _, block_num, *_, conf, text = word_data
32            block_num, conf = int(block_num), int(conf)
33
34            # handle line breaks
35            if block < block_num:
36                block = block_num
37                if self.words and self.words[-1].text != '\n':
38                    self.words.append(PredictedWord(0, '\n'))
39
40            # word predictions with low confidence will be filtered out
41            if conf >= conf_threshold:
42                self.words.append(PredictedWord(conf, text))
43
44        self.confidence = sum(word.confidence for word in self.words)
45
46        self.text = ' '.join(word.text for word in self.words)
47        # remove chars that are obviously ocr errors
48        table = str.maketrans('|', 'I', '<>{}[];`@#$%^*_=~\\')
49        self.text = self.text.translate(table).replace(' \n ', '\n').strip()
50
51    def is_similar_to(self, other: PredictedFrame, threshold=70) -> bool:
52        return fuzz.ratio(self.text, other.text) >= threshold
53
54
55class PredictedSubtitle:
56    frames: List[PredictedFrame]
57    text: str
58
59    def __init__(self, frames: List[PredictedFrame]):
60        self.frames = [f for f in frames if f.confidence > 0]
61
62        if self.frames:
63            self.text = max(self.frames, key=lambda f: f.confidence).text
64        else:
65            self.text = ''
66
67    @property
68    def index_start(self) -> int:
69        if self.frames:
70            return self.frames[0].index
71        return 0
72
73    @property
74    def index_end(self) -> int:
75        if self.frames:
76            return self.frames[-1].index
77        return 0
78
79    def is_similar_to(self, other: PredictedSubtitle, threshold=90) -> bool:
80        return fuzz.partial_ratio(self.text, other.text) >= threshold
81
82    def __repr__(self):
83        return '{} - {}. {}'.format(self.index_start, self.index_end, self.text)