docling_pdf/
layout.rs

1//! Layout detection via the RT-DETR (`docling-layout-heron`) model exported to
2//! ONNX, run with `ort`. A port of docling-ibm-models' `LayoutPredictor`:
3//! resize the page image to 640×640 and rescale to `[0,1]` (the heron processor
4//! has `do_normalize=false`), run the model, then RT-DETR
5//! `post_process_object_detection` (sigmoid → top-k over query×class →
6//! center-to-corners boxes scaled to the page).
7
8use image::imageops::FilterType;
9use image::RgbImage;
10use ort::session::Session;
11use ort::value::Tensor;
12
13/// The 17 canonical layout classes, indexed by the model's class id
14/// (`config.json` `id2label`).
15pub const LABELS: [&str; 17] = [
16    "caption",
17    "footnote",
18    "formula",
19    "list_item",
20    "page_footer",
21    "page_header",
22    "picture",
23    "section_header",
24    "table",
25    "text",
26    "title",
27    "document_index",
28    "code",
29    "checkbox_selected",
30    "checkbox_unselected",
31    "form",
32    "key_value_region",
33];
34
35/// One detected region, in page points (top-left origin).
36#[derive(Debug, Clone)]
37pub struct Region {
38    pub label: &'static str,
39    pub score: f32,
40    pub l: f32,
41    pub t: f32,
42    pub r: f32,
43    pub b: f32,
44}
45
46/// Base confidence threshold (docling-ibm-models `base_threshold`): the raw
47/// RT-DETR floor before docling's `LayoutPostprocessor` applies its stricter
48/// per-label thresholds ([`label_threshold`]).
49const THRESHOLD: f32 = 0.3;
50const SIDE: u32 = 640;
51
52/// Per-label confidence threshold, ported from docling's
53/// `LayoutPostprocessor.CONFIDENCE_THRESHOLDS`. The raw predictor keeps every
54/// detection above the 0.3 base; the postprocessor then drops a cluster whose
55/// score is below its label's threshold. Applying it here (equivalent, since
56/// every per-label threshold is ≥ the 0.3 base) keeps low-confidence pictures /
57/// tables / list-items out of the assembly, matching docling.
58pub fn label_threshold(label: &str) -> f32 {
59    match label {
60        "section_header"
61        | "title"
62        | "code"
63        | "checkbox_selected"
64        | "checkbox_unselected"
65        | "form"
66        | "key_value_region"
67        | "document_index" => 0.45,
68        // caption, footnote, formula, list_item, page_footer, page_header,
69        // picture, table, text — all 0.5 in docling.
70        _ => 0.5,
71    }
72}
73
74pub struct LayoutModel {
75    session: Session,
76}
77
78impl LayoutModel {
79    /// Load the ONNX model from `DOCLING_LAYOUT_ONNX`. Without the override,
80    /// prefers `models/layout_heron_int8.onnx` when present (the quantized
81    /// default; `DOCLING_RS_FP32=1` opts out), else `models/layout_heron.onnx`.
82    pub fn load() -> Result<Self, String> {
83        Self::load_with(crate::intra_threads())
84    }
85
86    /// Like [`load`](Self::load) but with an explicit intra-op thread count. A
87    /// parallel page-worker pool loads its helper models on a single thread each
88    /// and gets its speed-up from running pages concurrently instead.
89    pub fn load_with(intra: usize) -> Result<Self, String> {
90        let path = crate::model_path(
91            "DOCLING_LAYOUT_ONNX",
92            "models/layout_heron.onnx",
93            "models/layout_heron_int8.onnx",
94        );
95        let session = Session::builder()
96            .map_err(|e| format!("layout: builder: {e}"))?
97            // Let inference use the available cores (ort otherwise defaults low);
98            // a large PDF runs this model once per page.
99            .with_intra_threads(intra)
100            .map_err(|e| format!("layout: intra_threads: {e}"))?
101            .commit_from_file(&path)
102            .map_err(|e| format!("layout: load {path}: {e}"))?;
103        Ok(Self { session })
104    }
105
106    /// Detect layout regions on a page image. `page_w`/`page_h` are the page size
107    /// in points; returned boxes are in those coordinates.
108    pub fn predict(
109        &mut self,
110        img: &RgbImage,
111        page_w: f32,
112        page_h: f32,
113    ) -> Result<Vec<Region>, String> {
114        // Resize to 640×640 (RT-DETR ignores aspect ratio), rescale to [0,1],
115        // lay out as CHW.
116        let resized = image::imageops::resize(img, SIDE, SIDE, FilterType::Triangle);
117        let n = (SIDE * SIDE) as usize;
118        let mut data = vec![0f32; 3 * n];
119        for (i, px) in resized.pixels().enumerate() {
120            data[i] = px[0] as f32 / 255.0;
121            data[n + i] = px[1] as f32 / 255.0;
122            data[2 * n + i] = px[2] as f32 / 255.0;
123        }
124        let input = Tensor::from_array(([1usize, 3, SIDE as usize, SIDE as usize], data))
125            .map_err(|e| format!("layout: input tensor: {e}"))?;
126        let outputs = self
127            .session
128            .run(ort::inputs!["pixel_values" => input])
129            .map_err(|e| format!("layout: inference: {e}"))?;
130        let (lshape, logits) = outputs["logits"]
131            .try_extract_tensor::<f32>()
132            .map_err(|e| format!("layout: extract logits: {e}"))?;
133        let (_, boxes) = outputs["pred_boxes"]
134            .try_extract_tensor::<f32>()
135            .map_err(|e| format!("layout: extract boxes: {e}"))?;
136
137        let num_queries = lshape[1] as usize;
138        let num_classes = lshape[2] as usize;
139
140        // sigmoid over every (query, class); take the top `num_queries` scores.
141        let mut scored: Vec<(f32, usize)> = (0..num_queries * num_classes)
142            .map(|idx| (sigmoid(logits[idx]), idx))
143            .collect();
144        scored.sort_unstable_by(|a, b| b.0.total_cmp(&a.0));
145        scored.truncate(num_queries);
146
147        let mut regions = Vec::new();
148        for (score, idx) in scored {
149            if score <= THRESHOLD {
150                continue;
151            }
152            let label_id = idx % num_classes;
153            let q = idx / num_classes;
154            let cx = boxes[q * 4];
155            let cy = boxes[q * 4 + 1];
156            let w = boxes[q * 4 + 2];
157            let h = boxes[q * 4 + 3];
158            // center_to_corners, then scale normalized coords to page points.
159            let l = (cx - w / 2.0) * page_w;
160            let t = (cy - h / 2.0) * page_h;
161            let r = (cx + w / 2.0) * page_w;
162            let b = (cy + h / 2.0) * page_h;
163            regions.push(Region {
164                label: LABELS.get(label_id).copied().unwrap_or("text"),
165                score,
166                l,
167                t,
168                r,
169                b,
170            });
171        }
172        Ok(regions)
173    }
174}
175
176fn sigmoid(x: f32) -> f32 {
177    1.0 / (1.0 + (-x).exp())
178}
docling_pdf/layout.rs

docling_pdf/
layout.rs
