oar-ocr-core 0.6.3

Core types and predictors for oar-ocr
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173

use super::DBPostProcess;
use crate::processors::geometry::{BoundingBox, ScanlineBuffer};
use itertools::Itertools;
use rayon::prelude::*;

#[derive(Debug, Clone, Copy)]
struct Region {
    start_y: usize,
    end_y: usize,
    start_x: usize,
    end_x: usize,
}

impl Region {
    fn new(start_y: usize, end_y: usize, start_x: usize, end_x: usize) -> Self {
        Self {
            start_y,
            end_y,
            start_x,
            end_x,
        }
    }

    fn height(&self) -> usize {
        self.end_y - self.start_y
    }

    fn width(&self) -> usize {
        self.end_x - self.start_x
    }
}

impl DBPostProcess {
    /// Calculates the score of a bounding box using a fast approximation method.
    pub fn box_score_fast(&self, pred: &ndarray::ArrayView2<f32>, bbox: &BoundingBox) -> f32 {
        let height = pred.shape()[0];
        let width = pred.shape()[1];

        let (min_x, max_x) = bbox
            .points
            .iter()
            .map(|p| p.x)
            .minmax()
            .into_option()
            .unwrap_or((0.0, 0.0));
        let (min_y, max_y) = bbox
            .points
            .iter()
            .map(|p| p.y)
            .minmax()
            .into_option()
            .unwrap_or((0.0, 0.0));

        // Match PaddleX / OpenCV semantics:
        // xmin,ymin use floor; xmax,ymax use ceil before inclusive slicing.
        let min_x = min_x.floor().max(0.0).min(width as f32 - 1.0);
        let max_x = max_x.ceil().max(0.0).min(width as f32 - 1.0);
        let min_y = min_y.floor().max(0.0).min(height as f32 - 1.0);
        let max_y = max_y.ceil().max(0.0).min(height as f32 - 1.0);

        let start_y = min_y as usize;
        let end_y = max_y as usize + 1;
        let start_x = min_x as usize;
        let end_x = max_x as usize + 1;

        self.box_score_fast_contour(pred, bbox, start_y, end_y, start_x, end_x)
    }

    fn box_score_fast_contour(
        &self,
        pred: &ndarray::ArrayView2<f32>,
        bbox: &BoundingBox,
        start_y: usize,
        end_y: usize,
        start_x: usize,
        end_x: usize,
    ) -> f32 {
        let region = Region::new(start_y, end_y, start_x, end_x);
        self.box_score_fast_contour_with_policy(pred, bbox, region, None)
    }

    fn box_score_fast_contour_with_policy(
        &self,
        pred: &ndarray::ArrayView2<f32>,
        bbox: &BoundingBox,
        region: Region,
        policy: Option<&crate::core::config::ParallelPolicy>,
    ) -> f32 {
        let region_height = region.height();
        let region_width = region.width();

        let max_polygon_points = bbox.points.len();
        let mut scanline_buffer = ScanlineBuffer::new(max_polygon_points);

        let pixel_threshold = policy
            .map(|p| p.postprocess_pixel_threshold)
            .unwrap_or(8_000);

        if region_height * region_width < pixel_threshold {
            let mut total_score = 0.0;
            let mut total_pixels = 0;

            for y in region.start_y..region.end_y {
                let scanline_y = y as f32 + 0.5;
                let (line_score, line_pixels) = scanline_buffer.process_scanline(
                    scanline_y,
                    bbox,
                    region.start_x,
                    region.end_x,
                    pred,
                );
                total_score += line_score;
                total_pixels += line_pixels;
            }

            if total_pixels > 0 {
                total_score / total_pixels as f32
            } else {
                0.0
            }
        } else {
            let scanline_results: Vec<(f32, usize)> = (region.start_y..region.end_y)
                .into_par_iter()
                .map(|y| {
                    let scanline_y = y as f32 + 0.5;

                    let mut thread_buffer = ScanlineBuffer::new(max_polygon_points);
                    thread_buffer.process_scanline(
                        scanline_y,
                        bbox,
                        region.start_x,
                        region.end_x,
                        pred,
                    )
                })
                .collect();

            let total_score: f32 = scanline_results.iter().map(|(score, _)| score).sum();
            let total_pixels: usize = scanline_results.iter().map(|(_, pixels)| pixels).sum();

            if total_pixels > 0 {
                total_score / total_pixels as f32
            } else {
                0.0
            }
        }
    }

    pub(super) fn box_score_slow(
        &self,
        pred: &ndarray::ArrayView2<f32>,
        contour: &imageproc::contours::Contour<u32>,
    ) -> f32 {
        let mut total_score = 0.0;
        let mut pixel_count = 0;

        for point in &contour.points {
            let x = point.x as usize;
            let y = point.y as usize;

            if y < pred.shape()[0] && x < pred.shape()[1] {
                total_score += pred[[y, x]];
                pixel_count += 1;
            }
        }

        if pixel_count > 0 {
            total_score / pixel_count as f32
        } else {
            0.0
        }
    }
}