Struct OARStructureBuilder 

pub struct OARStructureBuilder { /* private fields */ }

High-level builder for document structure analysis pipelines.

This builder provides a fluent API for constructing document structure analysis pipelines with various components:

• Document preprocessing (optional): orientation detection and rectification
• Layout detection (required)
• Table classification (optional)
• Table cell detection (optional)
• Table structure recognition (optional)
• Formula recognition (optional)
• Seal text detection (optional)
• OCR integration (optional)

Example

use oar_ocr::oarocr::structure::OARStructureBuilder;

let structure = OARStructureBuilder::new("models/layout.onnx")
    .with_table_classification("models/table_cls.onnx")
    .with_table_cell_detection("models/table_cell.onnx", "wired")
    .with_table_structure_recognition("models/table_struct.onnx", "wired")
    .with_formula_recognition(
        "models/formula.onnx",
        "models/tokenizer.json",
        "pp_formulanet"
    )
    .build()?;

Implementations

impl OARStructureBuilder

pub fn new(layout_detection_model: impl Into<PathBuf>) -> Self

Creates a new structure builder with the required layout detection model.

Arguments

• layout_detection_model - Path to the layout detection model file

pub fn ort_session(self, config: OrtSessionConfig) -> Self

Sets the ONNX Runtime session configuration.

This configuration will be applied to all models in the pipeline.

pub fn layout_detection_config(self, config: LayoutDetectionConfig) -> Self

Sets the layout detection model configuration.

pub fn layout_model_name(self, name: impl Into<String>) -> Self

Overrides the built-in layout model preset used to configure preprocessing/postprocessing.

This is useful when the ONNX file name alone is not enough to infer the correct model family. Supported presets include:

• pp-doclayout_plus-l (default)
• pp-doclayout-s, pp-doclayout-m, pp-doclayout-l
• pp-docblocklayout
• picodet_layout_1x, picodet_layout_1x_table
• rt-detr-h_layout_3cls, rt-detr-h_layout_17cls

pub fn region_model_name(self, name: impl Into<String>) -> Self

Sets the region detection model name preset.

This is used to load the correct preprocessing/postprocessing for the region detection model. Supported presets: PP-DocBlockLayout.

pub fn wired_table_structure_model_name(self, name: impl Into<String>) -> Self

Sets the wired table structure model name preset.

Supported presets: SLANet, SLANeXt_wired.

pub fn wireless_table_structure_model_name( self, name: impl Into<String>, ) -> Self

Sets the wireless table structure model name preset.

Supported presets: SLANet_plus.

pub fn wired_table_cell_model_name(self, name: impl Into<String>) -> Self

Sets the wired table cell detection model name preset.

Supported presets: RT-DETR-L_wired_table_cell_det.

pub fn wireless_table_cell_model_name(self, name: impl Into<String>) -> Self

Sets the wireless table cell detection model name preset.

Supported presets: RT-DETR-L_wireless_table_cell_det.

pub fn text_detection_model_name(self, name: impl Into<String>) -> Self

Sets the text detection model name preset.

Supported presets: PP-OCRv5_mobile_det, PP-OCRv5_server_det.

pub fn text_recognition_model_name(self, name: impl Into<String>) -> Self

Sets the text recognition model name preset.

Supported presets: PP-OCRv5_mobile_rec, PP-OCRv5_server_rec.

pub fn image_batch_size(self, size: usize) -> Self

Sets the batch size for image-level processing.

Note: Currently not used in structure analysis as each image is processed individually.

pub fn region_batch_size(self, size: usize) -> Self

Sets the batch size for region-level processing (text recognition).

Controls how many text regions are processed together during OCR recognition. Larger values improve throughput but use more memory.

pub fn with_document_orientation(self, model_path: impl Into<PathBuf>) -> Self

Adds document orientation detection to the pipeline.

This component detects and corrects document rotation (0°, 90°, 180°, 270°). Should be run before other processing for best results.

pub fn with_document_rectification(self, model_path: impl Into<PathBuf>) -> Self

Adds document rectification to the pipeline.

This component corrects document distortion and perspective issues. Should be run after orientation detection if both are enabled.

pub fn with_region_detection(self, model_path: impl Into<PathBuf>) -> Self

Adds region detection to the pipeline (PP-DocBlockLayout).

This component detects document regions (columns, blocks) for hierarchical layout ordering. Region blocks provide grouping information for improved reading order within multi-column or complex layouts.

PP-StructureV3 Integration

When enabled, the pipeline uses region detection results to:

1. Group layout elements by their parent regions
2. Apply XY-cut ordering within each region
3. Order regions based on their relative positions

pub fn with_seal_text_detection(self, model_path: impl Into<PathBuf>) -> Self

Adds seal text detection to the pipeline.

This component detects circular/curved seal and stamp text regions. Seal regions will be included in the layout elements.

pub fn with_table_classification(self, model_path: impl Into<PathBuf>) -> Self

Adds table classification to the pipeline.

This component classifies tables as wired or wireless.

pub fn table_classification_config( self, config: TableClassificationConfig, ) -> Self

Sets the table classification configuration.

pub fn with_table_orientation(self, model_path: impl Into<PathBuf>) -> Self

Adds table orientation detection to the pipeline.

This component detects if tables are rotated (0°, 90°, 180°, 270°) and corrects them before structure recognition. Uses the same model as document orientation detection (PP-LCNet_x1_0_doc_ori).

Arguments

• model_path - Path to the orientation classification model (same as document orientation)

pub fn use_e2e_wired_table_rec(self, enabled: bool) -> Self

Sets whether to use end-to-end mode for wired table recognition.

When enabled, cell detection model is skipped and only the table structure recognition model’s cell output is used. When disabled, RT-DETR cell detection provides more precise cell bounding boxes.

Default: false (use cell detection for wired tables)

pub fn use_e2e_wireless_table_rec(self, enabled: bool) -> Self

Sets whether to use end-to-end mode for wireless table recognition.

When enabled, cell detection model is skipped and only the table structure recognition model’s cell output is used. When disabled, RT-DETR cell detection provides more precise cell bounding boxes.

Default: true (E2E mode for wireless tables)

pub fn use_wired_table_cells_trans_to_html(self, enabled: bool) -> Self

Enables PaddleX-compatible conversion of wired cell detections to HTML structure tokens.

When enabled, wired tables can derive structure tokens directly from detected cell boxes, which is useful for parity testing with PaddleX use_wired_table_cells_trans_to_html.

pub fn use_wireless_table_cells_trans_to_html(self, enabled: bool) -> Self

Enables PaddleX-compatible conversion of wireless cell detections to HTML structure tokens.

When enabled, wireless tables can derive structure tokens directly from detected cell boxes, which is useful for parity testing with PaddleX use_wireless_table_cells_trans_to_html.

pub fn with_table_cell_detection( self, model_path: impl Into<PathBuf>, cell_type: impl Into<String>, ) -> Self

Adds table cell detection to the pipeline.

Arguments

• model_path - Path to the table cell detection model
• cell_type - Type of cells to detect: “wired” or “wireless”

pub fn table_cell_detection_config( self, config: TableCellDetectionConfig, ) -> Self

Sets the table cell detection configuration.

pub fn with_table_structure_recognition( self, model_path: impl Into<PathBuf>, table_type: impl Into<String>, ) -> Self

Adds table structure recognition to the pipeline.

Arguments

• model_path - Path to the table structure recognition model
• table_type - Type of table structure: “wired” or “wireless”

This component recognizes the structure of tables and outputs HTML.

pub fn table_structure_dict_path(self, path: impl Into<PathBuf>) -> Self

Sets the dictionary path for table structure recognition.

The dictionary file should match the model type:

• table_structure_dict_ch.txt for Chinese
• table_structure_dict.txt for English
• table_master_structure_dict.txt for extended tags

pub fn table_structure_recognition_config( self, config: TableStructureRecognitionConfig, ) -> Self

Sets the table structure recognition configuration.

pub fn with_wired_table_structure(self, model_path: impl Into<PathBuf>) -> Self

Adds wired table structure recognition model.

When both wired and wireless models are configured along with table classification, the system automatically selects the appropriate model based on classification results.

pub fn with_wireless_table_structure( self, model_path: impl Into<PathBuf>, ) -> Self

Adds wireless table structure recognition model.

When both wired and wireless models are configured along with table classification, the system automatically selects the appropriate model based on classification results.

pub fn with_wired_table_cell_detection( self, model_path: impl Into<PathBuf>, ) -> Self

Adds wired table cell detection model.

When both wired and wireless models are configured along with table classification, the system automatically selects the appropriate model based on classification results.

pub fn with_wireless_table_cell_detection( self, model_path: impl Into<PathBuf>, ) -> Self

Adds wireless table cell detection model.

When both wired and wireless models are configured along with table classification, the system automatically selects the appropriate model based on classification results.

pub fn with_formula_recognition( self, model_path: impl Into<PathBuf>, tokenizer_path: impl Into<PathBuf>, model_type: impl Into<String>, ) -> Self

Adds formula recognition to the pipeline.

Arguments

• model_path - Path to the formula recognition model
• tokenizer_path - Path to the tokenizer JSON file
• model_type - Type of formula model: “pp_formulanet” or “unimernet”

This component recognizes mathematical formulas and outputs LaTeX.

pub fn formula_recognition_config( self, config: FormulaRecognitionConfig, ) -> Self

Sets the formula recognition configuration.

pub fn with_ocr( self, text_detection_model: impl Into<PathBuf>, text_recognition_model: impl Into<PathBuf>, character_dict_path: impl Into<PathBuf>, ) -> Self

Integrates OCR into the pipeline for text extraction.

Arguments

• text_detection_model - Path to the text detection model
• text_recognition_model - Path to the text recognition model
• character_dict_path - Path to the character dictionary file

pub fn with_text_line_orientation(self, model_path: impl Into<PathBuf>) -> Self

Adds text line orientation detection to the OCR pipeline.

This component detects whether text lines are upright (0°) or inverted (180°), which helps improve OCR accuracy for documents with mixed text orientations.

PP-StructureV3 Integration

When enabled, detected text lines are classified before recognition:

• Lines classified as 180° rotated are flipped before OCR
• This improves accuracy for documents scanned upside-down or with mixed orientations

pub fn text_detection_config(self, config: TextDetectionConfig) -> Self

Sets the text detection configuration.

pub fn text_recognition_config(self, config: TextRecognitionConfig) -> Self

Sets the text recognition configuration.

pub fn build(self) -> Result<OARStructure, OCRError>

Builds the structure analyzer runtime.

This method instantiates all adapters and returns a ready-to-use structure analyzer.

Trait Implementations

impl Clone for OARStructureBuilder

fn clone(&self) -> OARStructureBuilder

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for OARStructureBuilder

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.